1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
7 #include <linux/trace_recursion.h>
8 #include <linux/trace_events.h>
9 #include <linux/ring_buffer.h>
10 #include <linux/trace_clock.h>
11 #include <linux/sched/clock.h>
12 #include <linux/trace_seq.h>
13 #include <linux/spinlock.h>
14 #include <linux/irq_work.h>
15 #include <linux/security.h>
16 #include <linux/uaccess.h>
17 #include <linux/hardirq.h>
18 #include <linux/kthread.h> /* for self test */
19 #include <linux/module.h>
20 #include <linux/percpu.h>
21 #include <linux/mutex.h>
22 #include <linux/delay.h>
23 #include <linux/slab.h>
24 #include <linux/init.h>
25 #include <linux/hash.h>
26 #include <linux/list.h>
27 #include <linux/cpu.h>
28 #include <linux/oom.h>
30 #include <asm/local.h>
33 * The "absolute" timestamp in the buffer is only 59 bits.
34 * If a clock has the 5 MSBs set, it needs to be saved and
37 #define TS_MSB (0xf8ULL << 56)
38 #define ABS_TS_MASK (~TS_MSB)
40 static void update_pages_handler(struct work_struct *work);
43 * The ring buffer header is special. We must manually up keep it.
45 int ring_buffer_print_entry_header(struct trace_seq *s)
47 trace_seq_puts(s, "# compressed entry header\n");
48 trace_seq_puts(s, "\ttype_len : 5 bits\n");
49 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
50 trace_seq_puts(s, "\tarray : 32 bits\n");
51 trace_seq_putc(s, '\n');
52 trace_seq_printf(s, "\tpadding : type == %d\n",
53 RINGBUF_TYPE_PADDING);
54 trace_seq_printf(s, "\ttime_extend : type == %d\n",
55 RINGBUF_TYPE_TIME_EXTEND);
56 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
57 RINGBUF_TYPE_TIME_STAMP);
58 trace_seq_printf(s, "\tdata max type_len == %d\n",
59 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
61 return !trace_seq_has_overflowed(s);
65 * The ring buffer is made up of a list of pages. A separate list of pages is
66 * allocated for each CPU. A writer may only write to a buffer that is
67 * associated with the CPU it is currently executing on. A reader may read
68 * from any per cpu buffer.
70 * The reader is special. For each per cpu buffer, the reader has its own
71 * reader page. When a reader has read the entire reader page, this reader
72 * page is swapped with another page in the ring buffer.
74 * Now, as long as the writer is off the reader page, the reader can do what
75 * ever it wants with that page. The writer will never write to that page
76 * again (as long as it is out of the ring buffer).
78 * Here's some silly ASCII art.
81 * |reader| RING BUFFER
83 * +------+ +---+ +---+ +---+
92 * |reader| RING BUFFER
93 * |page |------------------v
94 * +------+ +---+ +---+ +---+
103 * |reader| RING BUFFER
104 * |page |------------------v
105 * +------+ +---+ +---+ +---+
107 * | +---+ +---+ +---+
110 * +------------------------------+
114 * |buffer| RING BUFFER
115 * |page |------------------v
116 * +------+ +---+ +---+ +---+
118 * | New +---+ +---+ +---+
121 * +------------------------------+
124 * After we make this swap, the reader can hand this page off to the splice
125 * code and be done with it. It can even allocate a new page if it needs to
126 * and swap that into the ring buffer.
128 * We will be using cmpxchg soon to make all this lockless.
132 /* Used for individual buffers (after the counter) */
133 #define RB_BUFFER_OFF (1 << 20)
135 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
137 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
138 #define RB_ALIGNMENT 4U
139 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
140 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
142 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
143 # define RB_FORCE_8BYTE_ALIGNMENT 0
144 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
146 # define RB_FORCE_8BYTE_ALIGNMENT 1
147 # define RB_ARCH_ALIGNMENT 8U
150 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
152 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
153 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
156 RB_LEN_TIME_EXTEND = 8,
157 RB_LEN_TIME_STAMP = 8,
160 #define skip_time_extend(event) \
161 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
163 #define extended_time(event) \
164 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
166 static inline bool rb_null_event(struct ring_buffer_event *event)
168 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
171 static void rb_event_set_padding(struct ring_buffer_event *event)
173 /* padding has a NULL time_delta */
174 event->type_len = RINGBUF_TYPE_PADDING;
175 event->time_delta = 0;
179 rb_event_data_length(struct ring_buffer_event *event)
184 length = event->type_len * RB_ALIGNMENT;
186 length = event->array[0];
187 return length + RB_EVNT_HDR_SIZE;
191 * Return the length of the given event. Will return
192 * the length of the time extend if the event is a
195 static inline unsigned
196 rb_event_length(struct ring_buffer_event *event)
198 switch (event->type_len) {
199 case RINGBUF_TYPE_PADDING:
200 if (rb_null_event(event))
203 return event->array[0] + RB_EVNT_HDR_SIZE;
205 case RINGBUF_TYPE_TIME_EXTEND:
206 return RB_LEN_TIME_EXTEND;
208 case RINGBUF_TYPE_TIME_STAMP:
209 return RB_LEN_TIME_STAMP;
211 case RINGBUF_TYPE_DATA:
212 return rb_event_data_length(event);
221 * Return total length of time extend and data,
222 * or just the event length for all other events.
224 static inline unsigned
225 rb_event_ts_length(struct ring_buffer_event *event)
229 if (extended_time(event)) {
230 /* time extends include the data event after it */
231 len = RB_LEN_TIME_EXTEND;
232 event = skip_time_extend(event);
234 return len + rb_event_length(event);
238 * ring_buffer_event_length - return the length of the event
239 * @event: the event to get the length of
241 * Returns the size of the data load of a data event.
242 * If the event is something other than a data event, it
243 * returns the size of the event itself. With the exception
244 * of a TIME EXTEND, where it still returns the size of the
245 * data load of the data event after it.
247 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
251 if (extended_time(event))
252 event = skip_time_extend(event);
254 length = rb_event_length(event);
255 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
257 length -= RB_EVNT_HDR_SIZE;
258 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
259 length -= sizeof(event->array[0]);
262 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
264 /* inline for ring buffer fast paths */
265 static __always_inline void *
266 rb_event_data(struct ring_buffer_event *event)
268 if (extended_time(event))
269 event = skip_time_extend(event);
270 WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
271 /* If length is in len field, then array[0] has the data */
273 return (void *)&event->array[0];
274 /* Otherwise length is in array[0] and array[1] has the data */
275 return (void *)&event->array[1];
279 * ring_buffer_event_data - return the data of the event
280 * @event: the event to get the data from
282 void *ring_buffer_event_data(struct ring_buffer_event *event)
284 return rb_event_data(event);
286 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
288 #define for_each_buffer_cpu(buffer, cpu) \
289 for_each_cpu(cpu, buffer->cpumask)
291 #define for_each_online_buffer_cpu(buffer, cpu) \
292 for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
295 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
296 #define TS_DELTA_TEST (~TS_MASK)
298 static u64 rb_event_time_stamp(struct ring_buffer_event *event)
302 ts = event->array[0];
304 ts += event->time_delta;
309 /* Flag when events were overwritten */
310 #define RB_MISSED_EVENTS (1 << 31)
311 /* Missed count stored at end */
312 #define RB_MISSED_STORED (1 << 30)
314 struct buffer_data_page {
315 u64 time_stamp; /* page time stamp */
316 local_t commit; /* write committed index */
317 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
321 * Note, the buffer_page list must be first. The buffer pages
322 * are allocated in cache lines, which means that each buffer
323 * page will be at the beginning of a cache line, and thus
324 * the least significant bits will be zero. We use this to
325 * add flags in the list struct pointers, to make the ring buffer
329 struct list_head list; /* list of buffer pages */
330 local_t write; /* index for next write */
331 unsigned read; /* index for next read */
332 local_t entries; /* entries on this page */
333 unsigned long real_end; /* real end of data */
334 struct buffer_data_page *page; /* Actual data page */
338 * The buffer page counters, write and entries, must be reset
339 * atomically when crossing page boundaries. To synchronize this
340 * update, two counters are inserted into the number. One is
341 * the actual counter for the write position or count on the page.
343 * The other is a counter of updaters. Before an update happens
344 * the update partition of the counter is incremented. This will
345 * allow the updater to update the counter atomically.
347 * The counter is 20 bits, and the state data is 12.
349 #define RB_WRITE_MASK 0xfffff
350 #define RB_WRITE_INTCNT (1 << 20)
352 static void rb_init_page(struct buffer_data_page *bpage)
354 local_set(&bpage->commit, 0);
357 static __always_inline unsigned int rb_page_commit(struct buffer_page *bpage)
359 return local_read(&bpage->page->commit);
362 static void free_buffer_page(struct buffer_page *bpage)
364 free_page((unsigned long)bpage->page);
369 * We need to fit the time_stamp delta into 27 bits.
371 static inline bool test_time_stamp(u64 delta)
373 return !!(delta & TS_DELTA_TEST);
376 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
378 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
379 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
381 int ring_buffer_print_page_header(struct trace_seq *s)
383 struct buffer_data_page field;
385 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
386 "offset:0;\tsize:%u;\tsigned:%u;\n",
387 (unsigned int)sizeof(field.time_stamp),
388 (unsigned int)is_signed_type(u64));
390 trace_seq_printf(s, "\tfield: local_t commit;\t"
391 "offset:%u;\tsize:%u;\tsigned:%u;\n",
392 (unsigned int)offsetof(typeof(field), commit),
393 (unsigned int)sizeof(field.commit),
394 (unsigned int)is_signed_type(long));
396 trace_seq_printf(s, "\tfield: int overwrite;\t"
397 "offset:%u;\tsize:%u;\tsigned:%u;\n",
398 (unsigned int)offsetof(typeof(field), commit),
400 (unsigned int)is_signed_type(long));
402 trace_seq_printf(s, "\tfield: char data;\t"
403 "offset:%u;\tsize:%u;\tsigned:%u;\n",
404 (unsigned int)offsetof(typeof(field), data),
405 (unsigned int)BUF_PAGE_SIZE,
406 (unsigned int)is_signed_type(char));
408 return !trace_seq_has_overflowed(s);
412 struct irq_work work;
413 wait_queue_head_t waiters;
414 wait_queue_head_t full_waiters;
416 bool waiters_pending;
417 bool full_waiters_pending;
422 * Structure to hold event state and handle nested events.
424 struct rb_event_info {
429 unsigned long length;
430 struct buffer_page *tail_page;
435 * Used for the add_timestamp
437 * EXTEND - wants a time extend
438 * ABSOLUTE - the buffer requests all events to have absolute time stamps
439 * FORCE - force a full time stamp.
442 RB_ADD_STAMP_NONE = 0,
443 RB_ADD_STAMP_EXTEND = BIT(1),
444 RB_ADD_STAMP_ABSOLUTE = BIT(2),
445 RB_ADD_STAMP_FORCE = BIT(3)
448 * Used for which event context the event is in.
455 * See trace_recursive_lock() comment below for more details.
466 #if BITS_PER_LONG == 32
470 /* To test on 64 bit machines */
475 struct rb_time_struct {
482 #include <asm/local64.h>
483 struct rb_time_struct {
487 typedef struct rb_time_struct rb_time_t;
492 * head_page == tail_page && head == tail then buffer is empty.
494 struct ring_buffer_per_cpu {
496 atomic_t record_disabled;
497 atomic_t resize_disabled;
498 struct trace_buffer *buffer;
499 raw_spinlock_t reader_lock; /* serialize readers */
500 arch_spinlock_t lock;
501 struct lock_class_key lock_key;
502 struct buffer_data_page *free_page;
503 unsigned long nr_pages;
504 unsigned int current_context;
505 struct list_head *pages;
506 struct buffer_page *head_page; /* read from head */
507 struct buffer_page *tail_page; /* write to tail */
508 struct buffer_page *commit_page; /* committed pages */
509 struct buffer_page *reader_page;
510 unsigned long lost_events;
511 unsigned long last_overrun;
513 local_t entries_bytes;
516 local_t commit_overrun;
517 local_t dropped_events;
520 local_t pages_touched;
523 long last_pages_touch;
524 size_t shortest_full;
526 unsigned long read_bytes;
527 rb_time_t write_stamp;
528 rb_time_t before_stamp;
529 u64 event_stamp[MAX_NEST];
531 /* pages removed since last reset */
532 unsigned long pages_removed;
533 /* ring buffer pages to update, > 0 to add, < 0 to remove */
534 long nr_pages_to_update;
535 struct list_head new_pages; /* new pages to add */
536 struct work_struct update_pages_work;
537 struct completion update_done;
539 struct rb_irq_work irq_work;
542 struct trace_buffer {
545 atomic_t record_disabled;
547 cpumask_var_t cpumask;
549 struct lock_class_key *reader_lock_key;
553 struct ring_buffer_per_cpu **buffers;
555 struct hlist_node node;
558 struct rb_irq_work irq_work;
562 struct ring_buffer_iter {
563 struct ring_buffer_per_cpu *cpu_buffer;
565 unsigned long next_event;
566 struct buffer_page *head_page;
567 struct buffer_page *cache_reader_page;
568 unsigned long cache_read;
569 unsigned long cache_pages_removed;
572 struct ring_buffer_event *event;
579 * On 32 bit machines, local64_t is very expensive. As the ring
580 * buffer doesn't need all the features of a true 64 bit atomic,
581 * on 32 bit, it uses these functions (64 still uses local64_t).
583 * For the ring buffer, 64 bit required operations for the time is
586 * - Reads may fail if it interrupted a modification of the time stamp.
587 * It will succeed if it did not interrupt another write even if
588 * the read itself is interrupted by a write.
589 * It returns whether it was successful or not.
591 * - Writes always succeed and will overwrite other writes and writes
592 * that were done by events interrupting the current write.
594 * - A write followed by a read of the same time stamp will always succeed,
595 * but may not contain the same value.
597 * - A cmpxchg will fail if it interrupted another write or cmpxchg.
598 * Other than that, it acts like a normal cmpxchg.
600 * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
601 * (bottom being the least significant 30 bits of the 60 bit time stamp).
603 * The two most significant bits of each half holds a 2 bit counter (0-3).
604 * Each update will increment this counter by one.
605 * When reading the top and bottom, if the two counter bits match then the
606 * top and bottom together make a valid 60 bit number.
608 #define RB_TIME_SHIFT 30
609 #define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
610 #define RB_TIME_MSB_SHIFT 60
612 static inline int rb_time_cnt(unsigned long val)
614 return (val >> RB_TIME_SHIFT) & 3;
617 static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
621 val = top & RB_TIME_VAL_MASK;
622 val <<= RB_TIME_SHIFT;
623 val |= bottom & RB_TIME_VAL_MASK;
628 static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
630 unsigned long top, bottom, msb;
634 * If the read is interrupted by a write, then the cnt will
635 * be different. Loop until both top and bottom have been read
636 * without interruption.
639 c = local_read(&t->cnt);
640 top = local_read(&t->top);
641 bottom = local_read(&t->bottom);
642 msb = local_read(&t->msb);
643 } while (c != local_read(&t->cnt));
645 *cnt = rb_time_cnt(top);
647 /* If top, msb or bottom counts don't match, this interrupted a write */
648 if (*cnt != rb_time_cnt(msb) || *cnt != rb_time_cnt(bottom))
651 /* The shift to msb will lose its cnt bits */
652 *ret = rb_time_val(top, bottom) | ((u64)msb << RB_TIME_MSB_SHIFT);
656 static bool rb_time_read(rb_time_t *t, u64 *ret)
660 return __rb_time_read(t, ret, &cnt);
663 static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
665 return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
668 static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom,
671 *top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
672 *bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
673 *msb = (unsigned long)(val >> RB_TIME_MSB_SHIFT);
676 static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
678 val = rb_time_val_cnt(val, cnt);
682 static void rb_time_set(rb_time_t *t, u64 val)
684 unsigned long cnt, top, bottom, msb;
686 rb_time_split(val, &top, &bottom, &msb);
688 /* Writes always succeed with a valid number even if it gets interrupted. */
690 cnt = local_inc_return(&t->cnt);
691 rb_time_val_set(&t->top, top, cnt);
692 rb_time_val_set(&t->bottom, bottom, cnt);
693 rb_time_val_set(&t->msb, val >> RB_TIME_MSB_SHIFT, cnt);
694 } while (cnt != local_read(&t->cnt));
698 rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
700 return local_try_cmpxchg(l, &expect, set);
703 static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
705 unsigned long cnt, top, bottom, msb;
706 unsigned long cnt2, top2, bottom2, msb2;
709 /* Any interruptions in this function should cause a failure */
710 cnt = local_read(&t->cnt);
712 /* The cmpxchg always fails if it interrupted an update */
713 if (!__rb_time_read(t, &val, &cnt2))
719 if ((cnt & 3) != cnt2)
724 rb_time_split(val, &top, &bottom, &msb);
725 msb = rb_time_val_cnt(msb, cnt);
726 top = rb_time_val_cnt(top, cnt);
727 bottom = rb_time_val_cnt(bottom, cnt);
729 rb_time_split(set, &top2, &bottom2, &msb2);
730 msb2 = rb_time_val_cnt(msb2, cnt);
731 top2 = rb_time_val_cnt(top2, cnt2);
732 bottom2 = rb_time_val_cnt(bottom2, cnt2);
734 if (!rb_time_read_cmpxchg(&t->cnt, cnt, cnt2))
736 if (!rb_time_read_cmpxchg(&t->msb, msb, msb2))
738 if (!rb_time_read_cmpxchg(&t->top, top, top2))
740 if (!rb_time_read_cmpxchg(&t->bottom, bottom, bottom2))
747 /* local64_t always succeeds */
749 static inline bool rb_time_read(rb_time_t *t, u64 *ret)
751 *ret = local64_read(&t->time);
754 static void rb_time_set(rb_time_t *t, u64 val)
756 local64_set(&t->time, val);
759 static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
761 return local64_try_cmpxchg(&t->time, &expect, set);
766 * Enable this to make sure that the event passed to
767 * ring_buffer_event_time_stamp() is not committed and also
768 * is on the buffer that it passed in.
770 //#define RB_VERIFY_EVENT
771 #ifdef RB_VERIFY_EVENT
772 static struct list_head *rb_list_head(struct list_head *list);
773 static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
776 struct buffer_page *page = cpu_buffer->commit_page;
777 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
778 struct list_head *next;
780 unsigned long addr = (unsigned long)event;
784 /* Make sure the event exists and is not committed yet */
786 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
788 commit = local_read(&page->page->commit);
789 write = local_read(&page->write);
790 if (addr >= (unsigned long)&page->page->data[commit] &&
791 addr < (unsigned long)&page->page->data[write])
794 next = rb_list_head(page->list.next);
795 page = list_entry(next, struct buffer_page, list);
800 static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
807 * The absolute time stamp drops the 5 MSBs and some clocks may
808 * require them. The rb_fix_abs_ts() will take a previous full
809 * time stamp, and add the 5 MSB of that time stamp on to the
810 * saved absolute time stamp. Then they are compared in case of
811 * the unlikely event that the latest time stamp incremented
814 static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
816 if (save_ts & TS_MSB) {
817 abs |= save_ts & TS_MSB;
818 /* Check for overflow */
819 if (unlikely(abs < save_ts))
825 static inline u64 rb_time_stamp(struct trace_buffer *buffer);
828 * ring_buffer_event_time_stamp - return the event's current time stamp
829 * @buffer: The buffer that the event is on
830 * @event: the event to get the time stamp of
832 * Note, this must be called after @event is reserved, and before it is
833 * committed to the ring buffer. And must be called from the same
834 * context where the event was reserved (normal, softirq, irq, etc).
836 * Returns the time stamp associated with the current event.
837 * If the event has an extended time stamp, then that is used as
838 * the time stamp to return.
839 * In the highly unlikely case that the event was nested more than
840 * the max nesting, then the write_stamp of the buffer is returned,
841 * otherwise current time is returned, but that really neither of
842 * the last two cases should ever happen.
844 u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
845 struct ring_buffer_event *event)
847 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
851 /* If the event includes an absolute time, then just use that */
852 if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
853 ts = rb_event_time_stamp(event);
854 return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
857 nest = local_read(&cpu_buffer->committing);
858 verify_event(cpu_buffer, event);
859 if (WARN_ON_ONCE(!nest))
862 /* Read the current saved nesting level time stamp */
863 if (likely(--nest < MAX_NEST))
864 return cpu_buffer->event_stamp[nest];
866 /* Shouldn't happen, warn if it does */
867 WARN_ONCE(1, "nest (%d) greater than max", nest);
870 /* Can only fail on 32 bit */
871 if (!rb_time_read(&cpu_buffer->write_stamp, &ts))
872 /* Screw it, just read the current time */
873 ts = rb_time_stamp(cpu_buffer->buffer);
879 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
880 * @buffer: The ring_buffer to get the number of pages from
881 * @cpu: The cpu of the ring_buffer to get the number of pages from
883 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
885 size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
887 return buffer->buffers[cpu]->nr_pages;
891 * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
892 * @buffer: The ring_buffer to get the number of pages from
893 * @cpu: The cpu of the ring_buffer to get the number of pages from
895 * Returns the number of pages that have content in the ring buffer.
897 size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
903 read = local_read(&buffer->buffers[cpu]->pages_read);
904 lost = local_read(&buffer->buffers[cpu]->pages_lost);
905 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
907 if (WARN_ON_ONCE(cnt < lost))
912 /* The reader can read an empty page, but not more than that */
914 WARN_ON_ONCE(read > cnt + 1);
921 static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
923 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
927 nr_pages = cpu_buffer->nr_pages;
928 if (!nr_pages || !full)
931 dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
933 return (dirty * 100) > (full * nr_pages);
937 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
939 * Schedules a delayed work to wake up any task that is blocked on the
940 * ring buffer waiters queue.
942 static void rb_wake_up_waiters(struct irq_work *work)
944 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
946 wake_up_all(&rbwork->waiters);
947 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
948 rbwork->wakeup_full = false;
949 rbwork->full_waiters_pending = false;
950 wake_up_all(&rbwork->full_waiters);
955 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
956 * @buffer: The ring buffer to wake waiters on
957 * @cpu: The CPU buffer to wake waiters on
959 * In the case of a file that represents a ring buffer is closing,
960 * it is prudent to wake up any waiters that are on this.
962 void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
964 struct ring_buffer_per_cpu *cpu_buffer;
965 struct rb_irq_work *rbwork;
970 if (cpu == RING_BUFFER_ALL_CPUS) {
972 /* Wake up individual ones too. One level recursion */
973 for_each_buffer_cpu(buffer, cpu)
974 ring_buffer_wake_waiters(buffer, cpu);
976 rbwork = &buffer->irq_work;
978 if (WARN_ON_ONCE(!buffer->buffers))
980 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
983 cpu_buffer = buffer->buffers[cpu];
984 /* The CPU buffer may not have been initialized yet */
987 rbwork = &cpu_buffer->irq_work;
990 rbwork->wait_index++;
991 /* make sure the waiters see the new index */
994 rb_wake_up_waiters(&rbwork->work);
998 * ring_buffer_wait - wait for input to the ring buffer
999 * @buffer: buffer to wait on
1000 * @cpu: the cpu buffer to wait on
1001 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
1003 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1004 * as data is added to any of the @buffer's cpu buffers. Otherwise
1005 * it will wait for data to be added to a specific cpu buffer.
1007 int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
1009 struct ring_buffer_per_cpu *cpu_buffer;
1011 struct rb_irq_work *work;
1016 * Depending on what the caller is waiting for, either any
1017 * data in any cpu buffer, or a specific buffer, put the
1018 * caller on the appropriate wait queue.
1020 if (cpu == RING_BUFFER_ALL_CPUS) {
1021 work = &buffer->irq_work;
1022 /* Full only makes sense on per cpu reads */
1025 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1027 cpu_buffer = buffer->buffers[cpu];
1028 work = &cpu_buffer->irq_work;
1031 wait_index = READ_ONCE(work->wait_index);
1035 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
1037 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
1040 * The events can happen in critical sections where
1041 * checking a work queue can cause deadlocks.
1042 * After adding a task to the queue, this flag is set
1043 * only to notify events to try to wake up the queue
1046 * We don't clear it even if the buffer is no longer
1047 * empty. The flag only causes the next event to run
1048 * irq_work to do the work queue wake up. The worse
1049 * that can happen if we race with !trace_empty() is that
1050 * an event will cause an irq_work to try to wake up
1053 * There's no reason to protect this flag either, as
1054 * the work queue and irq_work logic will do the necessary
1055 * synchronization for the wake ups. The only thing
1056 * that is necessary is that the wake up happens after
1057 * a task has been queued. It's OK for spurious wake ups.
1060 work->full_waiters_pending = true;
1062 work->waiters_pending = true;
1064 if (signal_pending(current)) {
1069 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
1072 if (cpu != RING_BUFFER_ALL_CPUS &&
1073 !ring_buffer_empty_cpu(buffer, cpu)) {
1074 unsigned long flags;
1081 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1082 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
1083 done = !pagebusy && full_hit(buffer, cpu, full);
1085 if (!cpu_buffer->shortest_full ||
1086 cpu_buffer->shortest_full > full)
1087 cpu_buffer->shortest_full = full;
1088 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1095 /* Make sure to see the new wait index */
1097 if (wait_index != work->wait_index)
1102 finish_wait(&work->full_waiters, &wait);
1104 finish_wait(&work->waiters, &wait);
1110 * ring_buffer_poll_wait - poll on buffer input
1111 * @buffer: buffer to wait on
1112 * @cpu: the cpu buffer to wait on
1113 * @filp: the file descriptor
1114 * @poll_table: The poll descriptor
1115 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
1117 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1118 * as data is added to any of the @buffer's cpu buffers. Otherwise
1119 * it will wait for data to be added to a specific cpu buffer.
1121 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
1124 __poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
1125 struct file *filp, poll_table *poll_table, int full)
1127 struct ring_buffer_per_cpu *cpu_buffer;
1128 struct rb_irq_work *work;
1130 if (cpu == RING_BUFFER_ALL_CPUS) {
1131 work = &buffer->irq_work;
1134 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1137 cpu_buffer = buffer->buffers[cpu];
1138 work = &cpu_buffer->irq_work;
1142 poll_wait(filp, &work->full_waiters, poll_table);
1143 work->full_waiters_pending = true;
1144 if (!cpu_buffer->shortest_full ||
1145 cpu_buffer->shortest_full > full)
1146 cpu_buffer->shortest_full = full;
1148 poll_wait(filp, &work->waiters, poll_table);
1149 work->waiters_pending = true;
1153 * There's a tight race between setting the waiters_pending and
1154 * checking if the ring buffer is empty. Once the waiters_pending bit
1155 * is set, the next event will wake the task up, but we can get stuck
1156 * if there's only a single event in.
1158 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1159 * but adding a memory barrier to all events will cause too much of a
1160 * performance hit in the fast path. We only need a memory barrier when
1161 * the buffer goes from empty to having content. But as this race is
1162 * extremely small, and it's not a problem if another event comes in, we
1163 * will fix it later.
1168 return full_hit(buffer, cpu, full) ? EPOLLIN | EPOLLRDNORM : 0;
1170 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1171 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1172 return EPOLLIN | EPOLLRDNORM;
1176 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
1177 #define RB_WARN_ON(b, cond) \
1179 int _____ret = unlikely(cond); \
1181 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1182 struct ring_buffer_per_cpu *__b = \
1184 atomic_inc(&__b->buffer->record_disabled); \
1186 atomic_inc(&b->record_disabled); \
1192 /* Up this if you want to test the TIME_EXTENTS and normalization */
1193 #define DEBUG_SHIFT 0
1195 static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1199 /* Skip retpolines :-( */
1200 if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1201 ts = trace_clock_local();
1203 ts = buffer->clock();
1205 /* shift to debug/test normalization and TIME_EXTENTS */
1206 return ts << DEBUG_SHIFT;
1209 u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1213 preempt_disable_notrace();
1214 time = rb_time_stamp(buffer);
1215 preempt_enable_notrace();
1219 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1221 void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1224 /* Just stupid testing the normalize function and deltas */
1225 *ts >>= DEBUG_SHIFT;
1227 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1230 * Making the ring buffer lockless makes things tricky.
1231 * Although writes only happen on the CPU that they are on,
1232 * and they only need to worry about interrupts. Reads can
1233 * happen on any CPU.
1235 * The reader page is always off the ring buffer, but when the
1236 * reader finishes with a page, it needs to swap its page with
1237 * a new one from the buffer. The reader needs to take from
1238 * the head (writes go to the tail). But if a writer is in overwrite
1239 * mode and wraps, it must push the head page forward.
1241 * Here lies the problem.
1243 * The reader must be careful to replace only the head page, and
1244 * not another one. As described at the top of the file in the
1245 * ASCII art, the reader sets its old page to point to the next
1246 * page after head. It then sets the page after head to point to
1247 * the old reader page. But if the writer moves the head page
1248 * during this operation, the reader could end up with the tail.
1250 * We use cmpxchg to help prevent this race. We also do something
1251 * special with the page before head. We set the LSB to 1.
1253 * When the writer must push the page forward, it will clear the
1254 * bit that points to the head page, move the head, and then set
1255 * the bit that points to the new head page.
1257 * We also don't want an interrupt coming in and moving the head
1258 * page on another writer. Thus we use the second LSB to catch
1261 * head->list->prev->next bit 1 bit 0
1264 * Points to head page 0 1
1267 * Note we can not trust the prev pointer of the head page, because:
1269 * +----+ +-----+ +-----+
1270 * | |------>| T |---X--->| N |
1272 * +----+ +-----+ +-----+
1275 * +----------| R |----------+ |
1279 * Key: ---X--> HEAD flag set in pointer
1284 * (see __rb_reserve_next() to see where this happens)
1286 * What the above shows is that the reader just swapped out
1287 * the reader page with a page in the buffer, but before it
1288 * could make the new header point back to the new page added
1289 * it was preempted by a writer. The writer moved forward onto
1290 * the new page added by the reader and is about to move forward
1293 * You can see, it is legitimate for the previous pointer of
1294 * the head (or any page) not to point back to itself. But only
1298 #define RB_PAGE_NORMAL 0UL
1299 #define RB_PAGE_HEAD 1UL
1300 #define RB_PAGE_UPDATE 2UL
1303 #define RB_FLAG_MASK 3UL
1305 /* PAGE_MOVED is not part of the mask */
1306 #define RB_PAGE_MOVED 4UL
1309 * rb_list_head - remove any bit
1311 static struct list_head *rb_list_head(struct list_head *list)
1313 unsigned long val = (unsigned long)list;
1315 return (struct list_head *)(val & ~RB_FLAG_MASK);
1319 * rb_is_head_page - test if the given page is the head page
1321 * Because the reader may move the head_page pointer, we can
1322 * not trust what the head page is (it may be pointing to
1323 * the reader page). But if the next page is a header page,
1324 * its flags will be non zero.
1327 rb_is_head_page(struct buffer_page *page, struct list_head *list)
1331 val = (unsigned long)list->next;
1333 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1334 return RB_PAGE_MOVED;
1336 return val & RB_FLAG_MASK;
1342 * The unique thing about the reader page, is that, if the
1343 * writer is ever on it, the previous pointer never points
1344 * back to the reader page.
1346 static bool rb_is_reader_page(struct buffer_page *page)
1348 struct list_head *list = page->list.prev;
1350 return rb_list_head(list->next) != &page->list;
1354 * rb_set_list_to_head - set a list_head to be pointing to head.
1356 static void rb_set_list_to_head(struct list_head *list)
1360 ptr = (unsigned long *)&list->next;
1361 *ptr |= RB_PAGE_HEAD;
1362 *ptr &= ~RB_PAGE_UPDATE;
1366 * rb_head_page_activate - sets up head page
1368 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1370 struct buffer_page *head;
1372 head = cpu_buffer->head_page;
1377 * Set the previous list pointer to have the HEAD flag.
1379 rb_set_list_to_head(head->list.prev);
1382 static void rb_list_head_clear(struct list_head *list)
1384 unsigned long *ptr = (unsigned long *)&list->next;
1386 *ptr &= ~RB_FLAG_MASK;
1390 * rb_head_page_deactivate - clears head page ptr (for free list)
1393 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1395 struct list_head *hd;
1397 /* Go through the whole list and clear any pointers found. */
1398 rb_list_head_clear(cpu_buffer->pages);
1400 list_for_each(hd, cpu_buffer->pages)
1401 rb_list_head_clear(hd);
1404 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1405 struct buffer_page *head,
1406 struct buffer_page *prev,
1407 int old_flag, int new_flag)
1409 struct list_head *list;
1410 unsigned long val = (unsigned long)&head->list;
1415 val &= ~RB_FLAG_MASK;
1417 ret = cmpxchg((unsigned long *)&list->next,
1418 val | old_flag, val | new_flag);
1420 /* check if the reader took the page */
1421 if ((ret & ~RB_FLAG_MASK) != val)
1422 return RB_PAGE_MOVED;
1424 return ret & RB_FLAG_MASK;
1427 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1428 struct buffer_page *head,
1429 struct buffer_page *prev,
1432 return rb_head_page_set(cpu_buffer, head, prev,
1433 old_flag, RB_PAGE_UPDATE);
1436 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1437 struct buffer_page *head,
1438 struct buffer_page *prev,
1441 return rb_head_page_set(cpu_buffer, head, prev,
1442 old_flag, RB_PAGE_HEAD);
1445 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1446 struct buffer_page *head,
1447 struct buffer_page *prev,
1450 return rb_head_page_set(cpu_buffer, head, prev,
1451 old_flag, RB_PAGE_NORMAL);
1454 static inline void rb_inc_page(struct buffer_page **bpage)
1456 struct list_head *p = rb_list_head((*bpage)->list.next);
1458 *bpage = list_entry(p, struct buffer_page, list);
1461 static struct buffer_page *
1462 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1464 struct buffer_page *head;
1465 struct buffer_page *page;
1466 struct list_head *list;
1469 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1473 list = cpu_buffer->pages;
1474 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1477 page = head = cpu_buffer->head_page;
1479 * It is possible that the writer moves the header behind
1480 * where we started, and we miss in one loop.
1481 * A second loop should grab the header, but we'll do
1482 * three loops just because I'm paranoid.
1484 for (i = 0; i < 3; i++) {
1486 if (rb_is_head_page(page, page->list.prev)) {
1487 cpu_buffer->head_page = page;
1491 } while (page != head);
1494 RB_WARN_ON(cpu_buffer, 1);
1499 static bool rb_head_page_replace(struct buffer_page *old,
1500 struct buffer_page *new)
1502 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1505 val = *ptr & ~RB_FLAG_MASK;
1506 val |= RB_PAGE_HEAD;
1508 return try_cmpxchg(ptr, &val, (unsigned long)&new->list);
1512 * rb_tail_page_update - move the tail page forward
1514 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1515 struct buffer_page *tail_page,
1516 struct buffer_page *next_page)
1518 unsigned long old_entries;
1519 unsigned long old_write;
1522 * The tail page now needs to be moved forward.
1524 * We need to reset the tail page, but without messing
1525 * with possible erasing of data brought in by interrupts
1526 * that have moved the tail page and are currently on it.
1528 * We add a counter to the write field to denote this.
1530 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1531 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1533 local_inc(&cpu_buffer->pages_touched);
1535 * Just make sure we have seen our old_write and synchronize
1536 * with any interrupts that come in.
1541 * If the tail page is still the same as what we think
1542 * it is, then it is up to us to update the tail
1545 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1546 /* Zero the write counter */
1547 unsigned long val = old_write & ~RB_WRITE_MASK;
1548 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1551 * This will only succeed if an interrupt did
1552 * not come in and change it. In which case, we
1553 * do not want to modify it.
1555 * We add (void) to let the compiler know that we do not care
1556 * about the return value of these functions. We use the
1557 * cmpxchg to only update if an interrupt did not already
1558 * do it for us. If the cmpxchg fails, we don't care.
1560 (void)local_cmpxchg(&next_page->write, old_write, val);
1561 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1564 * No need to worry about races with clearing out the commit.
1565 * it only can increment when a commit takes place. But that
1566 * only happens in the outer most nested commit.
1568 local_set(&next_page->page->commit, 0);
1570 /* Again, either we update tail_page or an interrupt does */
1571 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1575 static void rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1576 struct buffer_page *bpage)
1578 unsigned long val = (unsigned long)bpage;
1580 RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK);
1584 * rb_check_pages - integrity check of buffer pages
1585 * @cpu_buffer: CPU buffer with pages to test
1587 * As a safety measure we check to make sure the data pages have not
1590 static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1592 struct list_head *head = rb_list_head(cpu_buffer->pages);
1593 struct list_head *tmp;
1595 if (RB_WARN_ON(cpu_buffer,
1596 rb_list_head(rb_list_head(head->next)->prev) != head))
1599 if (RB_WARN_ON(cpu_buffer,
1600 rb_list_head(rb_list_head(head->prev)->next) != head))
1603 for (tmp = rb_list_head(head->next); tmp != head; tmp = rb_list_head(tmp->next)) {
1604 if (RB_WARN_ON(cpu_buffer,
1605 rb_list_head(rb_list_head(tmp->next)->prev) != tmp))
1608 if (RB_WARN_ON(cpu_buffer,
1609 rb_list_head(rb_list_head(tmp->prev)->next) != tmp))
1614 static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1615 long nr_pages, struct list_head *pages)
1617 struct buffer_page *bpage, *tmp;
1618 bool user_thread = current->mm != NULL;
1623 * Check if the available memory is there first.
1624 * Note, si_mem_available() only gives us a rough estimate of available
1625 * memory. It may not be accurate. But we don't care, we just want
1626 * to prevent doing any allocation when it is obvious that it is
1627 * not going to succeed.
1629 i = si_mem_available();
1634 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1635 * gracefully without invoking oom-killer and the system is not
1638 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1641 * If a user thread allocates too much, and si_mem_available()
1642 * reports there's enough memory, even though there is not.
1643 * Make sure the OOM killer kills this thread. This can happen
1644 * even with RETRY_MAYFAIL because another task may be doing
1645 * an allocation after this task has taken all memory.
1646 * This is the task the OOM killer needs to take out during this
1647 * loop, even if it was triggered by an allocation somewhere else.
1650 set_current_oom_origin();
1651 for (i = 0; i < nr_pages; i++) {
1654 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1655 mflags, cpu_to_node(cpu_buffer->cpu));
1659 rb_check_bpage(cpu_buffer, bpage);
1661 list_add(&bpage->list, pages);
1663 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags, 0);
1666 bpage->page = page_address(page);
1667 rb_init_page(bpage->page);
1669 if (user_thread && fatal_signal_pending(current))
1673 clear_current_oom_origin();
1678 list_for_each_entry_safe(bpage, tmp, pages, list) {
1679 list_del_init(&bpage->list);
1680 free_buffer_page(bpage);
1683 clear_current_oom_origin();
1688 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1689 unsigned long nr_pages)
1695 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1699 * The ring buffer page list is a circular list that does not
1700 * start and end with a list head. All page list items point to
1703 cpu_buffer->pages = pages.next;
1706 cpu_buffer->nr_pages = nr_pages;
1708 rb_check_pages(cpu_buffer);
1713 static struct ring_buffer_per_cpu *
1714 rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1716 struct ring_buffer_per_cpu *cpu_buffer;
1717 struct buffer_page *bpage;
1721 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1722 GFP_KERNEL, cpu_to_node(cpu));
1726 cpu_buffer->cpu = cpu;
1727 cpu_buffer->buffer = buffer;
1728 raw_spin_lock_init(&cpu_buffer->reader_lock);
1729 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1730 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1731 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1732 init_completion(&cpu_buffer->update_done);
1733 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1734 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1735 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1737 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1738 GFP_KERNEL, cpu_to_node(cpu));
1740 goto fail_free_buffer;
1742 rb_check_bpage(cpu_buffer, bpage);
1744 cpu_buffer->reader_page = bpage;
1745 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1747 goto fail_free_reader;
1748 bpage->page = page_address(page);
1749 rb_init_page(bpage->page);
1751 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1752 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1754 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1756 goto fail_free_reader;
1758 cpu_buffer->head_page
1759 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1760 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1762 rb_head_page_activate(cpu_buffer);
1767 free_buffer_page(cpu_buffer->reader_page);
1774 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1776 struct list_head *head = cpu_buffer->pages;
1777 struct buffer_page *bpage, *tmp;
1779 irq_work_sync(&cpu_buffer->irq_work.work);
1781 free_buffer_page(cpu_buffer->reader_page);
1784 rb_head_page_deactivate(cpu_buffer);
1786 list_for_each_entry_safe(bpage, tmp, head, list) {
1787 list_del_init(&bpage->list);
1788 free_buffer_page(bpage);
1790 bpage = list_entry(head, struct buffer_page, list);
1791 free_buffer_page(bpage);
1794 free_page((unsigned long)cpu_buffer->free_page);
1800 * __ring_buffer_alloc - allocate a new ring_buffer
1801 * @size: the size in bytes per cpu that is needed.
1802 * @flags: attributes to set for the ring buffer.
1803 * @key: ring buffer reader_lock_key.
1805 * Currently the only flag that is available is the RB_FL_OVERWRITE
1806 * flag. This flag means that the buffer will overwrite old data
1807 * when the buffer wraps. If this flag is not set, the buffer will
1808 * drop data when the tail hits the head.
1810 struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1811 struct lock_class_key *key)
1813 struct trace_buffer *buffer;
1819 /* keep it in its own cache line */
1820 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1825 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1826 goto fail_free_buffer;
1828 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1829 buffer->flags = flags;
1830 buffer->clock = trace_clock_local;
1831 buffer->reader_lock_key = key;
1833 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1834 init_waitqueue_head(&buffer->irq_work.waiters);
1836 /* need at least two pages */
1840 buffer->cpus = nr_cpu_ids;
1842 bsize = sizeof(void *) * nr_cpu_ids;
1843 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1845 if (!buffer->buffers)
1846 goto fail_free_cpumask;
1848 cpu = raw_smp_processor_id();
1849 cpumask_set_cpu(cpu, buffer->cpumask);
1850 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1851 if (!buffer->buffers[cpu])
1852 goto fail_free_buffers;
1854 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1856 goto fail_free_buffers;
1858 mutex_init(&buffer->mutex);
1863 for_each_buffer_cpu(buffer, cpu) {
1864 if (buffer->buffers[cpu])
1865 rb_free_cpu_buffer(buffer->buffers[cpu]);
1867 kfree(buffer->buffers);
1870 free_cpumask_var(buffer->cpumask);
1876 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1879 * ring_buffer_free - free a ring buffer.
1880 * @buffer: the buffer to free.
1883 ring_buffer_free(struct trace_buffer *buffer)
1887 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1889 irq_work_sync(&buffer->irq_work.work);
1891 for_each_buffer_cpu(buffer, cpu)
1892 rb_free_cpu_buffer(buffer->buffers[cpu]);
1894 kfree(buffer->buffers);
1895 free_cpumask_var(buffer->cpumask);
1899 EXPORT_SYMBOL_GPL(ring_buffer_free);
1901 void ring_buffer_set_clock(struct trace_buffer *buffer,
1904 buffer->clock = clock;
1907 void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1909 buffer->time_stamp_abs = abs;
1912 bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1914 return buffer->time_stamp_abs;
1917 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1919 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1921 return local_read(&bpage->entries) & RB_WRITE_MASK;
1924 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1926 return local_read(&bpage->write) & RB_WRITE_MASK;
1930 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1932 struct list_head *tail_page, *to_remove, *next_page;
1933 struct buffer_page *to_remove_page, *tmp_iter_page;
1934 struct buffer_page *last_page, *first_page;
1935 unsigned long nr_removed;
1936 unsigned long head_bit;
1941 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1942 atomic_inc(&cpu_buffer->record_disabled);
1944 * We don't race with the readers since we have acquired the reader
1945 * lock. We also don't race with writers after disabling recording.
1946 * This makes it easy to figure out the first and the last page to be
1947 * removed from the list. We unlink all the pages in between including
1948 * the first and last pages. This is done in a busy loop so that we
1949 * lose the least number of traces.
1950 * The pages are freed after we restart recording and unlock readers.
1952 tail_page = &cpu_buffer->tail_page->list;
1955 * tail page might be on reader page, we remove the next page
1956 * from the ring buffer
1958 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1959 tail_page = rb_list_head(tail_page->next);
1960 to_remove = tail_page;
1962 /* start of pages to remove */
1963 first_page = list_entry(rb_list_head(to_remove->next),
1964 struct buffer_page, list);
1966 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1967 to_remove = rb_list_head(to_remove)->next;
1968 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1970 /* Read iterators need to reset themselves when some pages removed */
1971 cpu_buffer->pages_removed += nr_removed;
1973 next_page = rb_list_head(to_remove)->next;
1976 * Now we remove all pages between tail_page and next_page.
1977 * Make sure that we have head_bit value preserved for the
1980 tail_page->next = (struct list_head *)((unsigned long)next_page |
1982 next_page = rb_list_head(next_page);
1983 next_page->prev = tail_page;
1985 /* make sure pages points to a valid page in the ring buffer */
1986 cpu_buffer->pages = next_page;
1988 /* update head page */
1990 cpu_buffer->head_page = list_entry(next_page,
1991 struct buffer_page, list);
1993 /* pages are removed, resume tracing and then free the pages */
1994 atomic_dec(&cpu_buffer->record_disabled);
1995 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1997 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1999 /* last buffer page to remove */
2000 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
2002 tmp_iter_page = first_page;
2007 to_remove_page = tmp_iter_page;
2008 rb_inc_page(&tmp_iter_page);
2010 /* update the counters */
2011 page_entries = rb_page_entries(to_remove_page);
2014 * If something was added to this page, it was full
2015 * since it is not the tail page. So we deduct the
2016 * bytes consumed in ring buffer from here.
2017 * Increment overrun to account for the lost events.
2019 local_add(page_entries, &cpu_buffer->overrun);
2020 local_sub(rb_page_commit(to_remove_page), &cpu_buffer->entries_bytes);
2021 local_inc(&cpu_buffer->pages_lost);
2025 * We have already removed references to this list item, just
2026 * free up the buffer_page and its page
2028 free_buffer_page(to_remove_page);
2031 } while (to_remove_page != last_page);
2033 RB_WARN_ON(cpu_buffer, nr_removed);
2035 return nr_removed == 0;
2039 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
2041 struct list_head *pages = &cpu_buffer->new_pages;
2042 unsigned long flags;
2046 /* Can be called at early boot up, where interrupts must not been enabled */
2047 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2049 * We are holding the reader lock, so the reader page won't be swapped
2050 * in the ring buffer. Now we are racing with the writer trying to
2051 * move head page and the tail page.
2052 * We are going to adapt the reader page update process where:
2053 * 1. We first splice the start and end of list of new pages between
2054 * the head page and its previous page.
2055 * 2. We cmpxchg the prev_page->next to point from head page to the
2056 * start of new pages list.
2057 * 3. Finally, we update the head->prev to the end of new list.
2059 * We will try this process 10 times, to make sure that we don't keep
2065 struct list_head *head_page, *prev_page, *r;
2066 struct list_head *last_page, *first_page;
2067 struct list_head *head_page_with_bit;
2068 struct buffer_page *hpage = rb_set_head_page(cpu_buffer);
2072 head_page = &hpage->list;
2073 prev_page = head_page->prev;
2075 first_page = pages->next;
2076 last_page = pages->prev;
2078 head_page_with_bit = (struct list_head *)
2079 ((unsigned long)head_page | RB_PAGE_HEAD);
2081 last_page->next = head_page_with_bit;
2082 first_page->prev = prev_page;
2084 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
2086 if (r == head_page_with_bit) {
2088 * yay, we replaced the page pointer to our new list,
2089 * now, we just have to update to head page's prev
2090 * pointer to point to end of list
2092 head_page->prev = last_page;
2099 INIT_LIST_HEAD(pages);
2101 * If we weren't successful in adding in new pages, warn and stop
2104 RB_WARN_ON(cpu_buffer, !success);
2105 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2107 /* free pages if they weren't inserted */
2109 struct buffer_page *bpage, *tmp;
2110 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2112 list_del_init(&bpage->list);
2113 free_buffer_page(bpage);
2119 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2123 if (cpu_buffer->nr_pages_to_update > 0)
2124 success = rb_insert_pages(cpu_buffer);
2126 success = rb_remove_pages(cpu_buffer,
2127 -cpu_buffer->nr_pages_to_update);
2130 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2133 static void update_pages_handler(struct work_struct *work)
2135 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2136 struct ring_buffer_per_cpu, update_pages_work);
2137 rb_update_pages(cpu_buffer);
2138 complete(&cpu_buffer->update_done);
2142 * ring_buffer_resize - resize the ring buffer
2143 * @buffer: the buffer to resize.
2144 * @size: the new size.
2145 * @cpu_id: the cpu buffer to resize
2147 * Minimum size is 2 * BUF_PAGE_SIZE.
2149 * Returns 0 on success and < 0 on failure.
2151 int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2154 struct ring_buffer_per_cpu *cpu_buffer;
2155 unsigned long nr_pages;
2159 * Always succeed at resizing a non-existent buffer:
2164 /* Make sure the requested buffer exists */
2165 if (cpu_id != RING_BUFFER_ALL_CPUS &&
2166 !cpumask_test_cpu(cpu_id, buffer->cpumask))
2169 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
2171 /* we need a minimum of two pages */
2175 /* prevent another thread from changing buffer sizes */
2176 mutex_lock(&buffer->mutex);
2177 atomic_inc(&buffer->resizing);
2179 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2181 * Don't succeed if resizing is disabled, as a reader might be
2182 * manipulating the ring buffer and is expecting a sane state while
2185 for_each_buffer_cpu(buffer, cpu) {
2186 cpu_buffer = buffer->buffers[cpu];
2187 if (atomic_read(&cpu_buffer->resize_disabled)) {
2189 goto out_err_unlock;
2193 /* calculate the pages to update */
2194 for_each_buffer_cpu(buffer, cpu) {
2195 cpu_buffer = buffer->buffers[cpu];
2197 cpu_buffer->nr_pages_to_update = nr_pages -
2198 cpu_buffer->nr_pages;
2200 * nothing more to do for removing pages or no update
2202 if (cpu_buffer->nr_pages_to_update <= 0)
2205 * to add pages, make sure all new pages can be
2206 * allocated without receiving ENOMEM
2208 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2209 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2210 &cpu_buffer->new_pages)) {
2211 /* not enough memory for new pages */
2221 * Fire off all the required work handlers
2222 * We can't schedule on offline CPUs, but it's not necessary
2223 * since we can change their buffer sizes without any race.
2225 for_each_buffer_cpu(buffer, cpu) {
2226 cpu_buffer = buffer->buffers[cpu];
2227 if (!cpu_buffer->nr_pages_to_update)
2230 /* Can't run something on an offline CPU. */
2231 if (!cpu_online(cpu)) {
2232 rb_update_pages(cpu_buffer);
2233 cpu_buffer->nr_pages_to_update = 0;
2235 /* Run directly if possible. */
2237 if (cpu != smp_processor_id()) {
2239 schedule_work_on(cpu,
2240 &cpu_buffer->update_pages_work);
2242 update_pages_handler(&cpu_buffer->update_pages_work);
2248 /* wait for all the updates to complete */
2249 for_each_buffer_cpu(buffer, cpu) {
2250 cpu_buffer = buffer->buffers[cpu];
2251 if (!cpu_buffer->nr_pages_to_update)
2254 if (cpu_online(cpu))
2255 wait_for_completion(&cpu_buffer->update_done);
2256 cpu_buffer->nr_pages_to_update = 0;
2261 cpu_buffer = buffer->buffers[cpu_id];
2263 if (nr_pages == cpu_buffer->nr_pages)
2267 * Don't succeed if resizing is disabled, as a reader might be
2268 * manipulating the ring buffer and is expecting a sane state while
2271 if (atomic_read(&cpu_buffer->resize_disabled)) {
2273 goto out_err_unlock;
2276 cpu_buffer->nr_pages_to_update = nr_pages -
2277 cpu_buffer->nr_pages;
2279 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2280 if (cpu_buffer->nr_pages_to_update > 0 &&
2281 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2282 &cpu_buffer->new_pages)) {
2289 /* Can't run something on an offline CPU. */
2290 if (!cpu_online(cpu_id))
2291 rb_update_pages(cpu_buffer);
2293 /* Run directly if possible. */
2295 if (cpu_id == smp_processor_id()) {
2296 rb_update_pages(cpu_buffer);
2300 schedule_work_on(cpu_id,
2301 &cpu_buffer->update_pages_work);
2302 wait_for_completion(&cpu_buffer->update_done);
2306 cpu_buffer->nr_pages_to_update = 0;
2312 * The ring buffer resize can happen with the ring buffer
2313 * enabled, so that the update disturbs the tracing as little
2314 * as possible. But if the buffer is disabled, we do not need
2315 * to worry about that, and we can take the time to verify
2316 * that the buffer is not corrupt.
2318 if (atomic_read(&buffer->record_disabled)) {
2319 atomic_inc(&buffer->record_disabled);
2321 * Even though the buffer was disabled, we must make sure
2322 * that it is truly disabled before calling rb_check_pages.
2323 * There could have been a race between checking
2324 * record_disable and incrementing it.
2327 for_each_buffer_cpu(buffer, cpu) {
2328 cpu_buffer = buffer->buffers[cpu];
2329 rb_check_pages(cpu_buffer);
2331 atomic_dec(&buffer->record_disabled);
2334 atomic_dec(&buffer->resizing);
2335 mutex_unlock(&buffer->mutex);
2339 for_each_buffer_cpu(buffer, cpu) {
2340 struct buffer_page *bpage, *tmp;
2342 cpu_buffer = buffer->buffers[cpu];
2343 cpu_buffer->nr_pages_to_update = 0;
2345 if (list_empty(&cpu_buffer->new_pages))
2348 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2350 list_del_init(&bpage->list);
2351 free_buffer_page(bpage);
2355 atomic_dec(&buffer->resizing);
2356 mutex_unlock(&buffer->mutex);
2359 EXPORT_SYMBOL_GPL(ring_buffer_resize);
2361 void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2363 mutex_lock(&buffer->mutex);
2365 buffer->flags |= RB_FL_OVERWRITE;
2367 buffer->flags &= ~RB_FL_OVERWRITE;
2368 mutex_unlock(&buffer->mutex);
2370 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2372 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2374 return bpage->page->data + index;
2377 static __always_inline struct ring_buffer_event *
2378 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2380 return __rb_page_index(cpu_buffer->reader_page,
2381 cpu_buffer->reader_page->read);
2384 static struct ring_buffer_event *
2385 rb_iter_head_event(struct ring_buffer_iter *iter)
2387 struct ring_buffer_event *event;
2388 struct buffer_page *iter_head_page = iter->head_page;
2389 unsigned long commit;
2392 if (iter->head != iter->next_event)
2396 * When the writer goes across pages, it issues a cmpxchg which
2397 * is a mb(), which will synchronize with the rmb here.
2398 * (see rb_tail_page_update() and __rb_reserve_next())
2400 commit = rb_page_commit(iter_head_page);
2403 /* An event needs to be at least 8 bytes in size */
2404 if (iter->head > commit - 8)
2407 event = __rb_page_index(iter_head_page, iter->head);
2408 length = rb_event_length(event);
2411 * READ_ONCE() doesn't work on functions and we don't want the
2412 * compiler doing any crazy optimizations with length.
2416 if ((iter->head + length) > commit || length > BUF_PAGE_SIZE)
2417 /* Writer corrupted the read? */
2420 memcpy(iter->event, event, length);
2422 * If the page stamp is still the same after this rmb() then the
2423 * event was safely copied without the writer entering the page.
2427 /* Make sure the page didn't change since we read this */
2428 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2429 commit > rb_page_commit(iter_head_page))
2432 iter->next_event = iter->head + length;
2435 /* Reset to the beginning */
2436 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2438 iter->next_event = 0;
2439 iter->missed_events = 1;
2443 /* Size is determined by what has been committed */
2444 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2446 return rb_page_commit(bpage);
2449 static __always_inline unsigned
2450 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2452 return rb_page_commit(cpu_buffer->commit_page);
2455 static __always_inline unsigned
2456 rb_event_index(struct ring_buffer_event *event)
2458 unsigned long addr = (unsigned long)event;
2460 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2463 static void rb_inc_iter(struct ring_buffer_iter *iter)
2465 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2468 * The iterator could be on the reader page (it starts there).
2469 * But the head could have moved, since the reader was
2470 * found. Check for this case and assign the iterator
2471 * to the head page instead of next.
2473 if (iter->head_page == cpu_buffer->reader_page)
2474 iter->head_page = rb_set_head_page(cpu_buffer);
2476 rb_inc_page(&iter->head_page);
2478 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2480 iter->next_event = 0;
2484 * rb_handle_head_page - writer hit the head page
2486 * Returns: +1 to retry page
2491 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2492 struct buffer_page *tail_page,
2493 struct buffer_page *next_page)
2495 struct buffer_page *new_head;
2500 entries = rb_page_entries(next_page);
2503 * The hard part is here. We need to move the head
2504 * forward, and protect against both readers on
2505 * other CPUs and writers coming in via interrupts.
2507 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2511 * type can be one of four:
2512 * NORMAL - an interrupt already moved it for us
2513 * HEAD - we are the first to get here.
2514 * UPDATE - we are the interrupt interrupting
2516 * MOVED - a reader on another CPU moved the next
2517 * pointer to its reader page. Give up
2524 * We changed the head to UPDATE, thus
2525 * it is our responsibility to update
2528 local_add(entries, &cpu_buffer->overrun);
2529 local_sub(rb_page_commit(next_page), &cpu_buffer->entries_bytes);
2530 local_inc(&cpu_buffer->pages_lost);
2533 * The entries will be zeroed out when we move the
2537 /* still more to do */
2540 case RB_PAGE_UPDATE:
2542 * This is an interrupt that interrupt the
2543 * previous update. Still more to do.
2546 case RB_PAGE_NORMAL:
2548 * An interrupt came in before the update
2549 * and processed this for us.
2550 * Nothing left to do.
2555 * The reader is on another CPU and just did
2556 * a swap with our next_page.
2561 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2566 * Now that we are here, the old head pointer is
2567 * set to UPDATE. This will keep the reader from
2568 * swapping the head page with the reader page.
2569 * The reader (on another CPU) will spin till
2572 * We just need to protect against interrupts
2573 * doing the job. We will set the next pointer
2574 * to HEAD. After that, we set the old pointer
2575 * to NORMAL, but only if it was HEAD before.
2576 * otherwise we are an interrupt, and only
2577 * want the outer most commit to reset it.
2579 new_head = next_page;
2580 rb_inc_page(&new_head);
2582 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2586 * Valid returns are:
2587 * HEAD - an interrupt came in and already set it.
2588 * NORMAL - One of two things:
2589 * 1) We really set it.
2590 * 2) A bunch of interrupts came in and moved
2591 * the page forward again.
2595 case RB_PAGE_NORMAL:
2599 RB_WARN_ON(cpu_buffer, 1);
2604 * It is possible that an interrupt came in,
2605 * set the head up, then more interrupts came in
2606 * and moved it again. When we get back here,
2607 * the page would have been set to NORMAL but we
2608 * just set it back to HEAD.
2610 * How do you detect this? Well, if that happened
2611 * the tail page would have moved.
2613 if (ret == RB_PAGE_NORMAL) {
2614 struct buffer_page *buffer_tail_page;
2616 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2618 * If the tail had moved passed next, then we need
2619 * to reset the pointer.
2621 if (buffer_tail_page != tail_page &&
2622 buffer_tail_page != next_page)
2623 rb_head_page_set_normal(cpu_buffer, new_head,
2629 * If this was the outer most commit (the one that
2630 * changed the original pointer from HEAD to UPDATE),
2631 * then it is up to us to reset it to NORMAL.
2633 if (type == RB_PAGE_HEAD) {
2634 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2637 if (RB_WARN_ON(cpu_buffer,
2638 ret != RB_PAGE_UPDATE))
2646 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2647 unsigned long tail, struct rb_event_info *info)
2649 struct buffer_page *tail_page = info->tail_page;
2650 struct ring_buffer_event *event;
2651 unsigned long length = info->length;
2654 * Only the event that crossed the page boundary
2655 * must fill the old tail_page with padding.
2657 if (tail >= BUF_PAGE_SIZE) {
2659 * If the page was filled, then we still need
2660 * to update the real_end. Reset it to zero
2661 * and the reader will ignore it.
2663 if (tail == BUF_PAGE_SIZE)
2664 tail_page->real_end = 0;
2666 local_sub(length, &tail_page->write);
2670 event = __rb_page_index(tail_page, tail);
2673 * Save the original length to the meta data.
2674 * This will be used by the reader to add lost event
2677 tail_page->real_end = tail;
2680 * If this event is bigger than the minimum size, then
2681 * we need to be careful that we don't subtract the
2682 * write counter enough to allow another writer to slip
2684 * We put in a discarded commit instead, to make sure
2685 * that this space is not used again, and this space will
2686 * not be accounted into 'entries_bytes'.
2688 * If we are less than the minimum size, we don't need to
2691 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2692 /* No room for any events */
2694 /* Mark the rest of the page with padding */
2695 rb_event_set_padding(event);
2697 /* Make sure the padding is visible before the write update */
2700 /* Set the write back to the previous setting */
2701 local_sub(length, &tail_page->write);
2705 /* Put in a discarded event */
2706 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2707 event->type_len = RINGBUF_TYPE_PADDING;
2708 /* time delta must be non zero */
2709 event->time_delta = 1;
2711 /* account for padding bytes */
2712 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2714 /* Make sure the padding is visible before the tail_page->write update */
2717 /* Set write to end of buffer */
2718 length = (tail + length) - BUF_PAGE_SIZE;
2719 local_sub(length, &tail_page->write);
2722 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2725 * This is the slow path, force gcc not to inline it.
2727 static noinline struct ring_buffer_event *
2728 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2729 unsigned long tail, struct rb_event_info *info)
2731 struct buffer_page *tail_page = info->tail_page;
2732 struct buffer_page *commit_page = cpu_buffer->commit_page;
2733 struct trace_buffer *buffer = cpu_buffer->buffer;
2734 struct buffer_page *next_page;
2737 next_page = tail_page;
2739 rb_inc_page(&next_page);
2742 * If for some reason, we had an interrupt storm that made
2743 * it all the way around the buffer, bail, and warn
2746 if (unlikely(next_page == commit_page)) {
2747 local_inc(&cpu_buffer->commit_overrun);
2752 * This is where the fun begins!
2754 * We are fighting against races between a reader that
2755 * could be on another CPU trying to swap its reader
2756 * page with the buffer head.
2758 * We are also fighting against interrupts coming in and
2759 * moving the head or tail on us as well.
2761 * If the next page is the head page then we have filled
2762 * the buffer, unless the commit page is still on the
2765 if (rb_is_head_page(next_page, &tail_page->list)) {
2768 * If the commit is not on the reader page, then
2769 * move the header page.
2771 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2773 * If we are not in overwrite mode,
2774 * this is easy, just stop here.
2776 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2777 local_inc(&cpu_buffer->dropped_events);
2781 ret = rb_handle_head_page(cpu_buffer,
2790 * We need to be careful here too. The
2791 * commit page could still be on the reader
2792 * page. We could have a small buffer, and
2793 * have filled up the buffer with events
2794 * from interrupts and such, and wrapped.
2796 * Note, if the tail page is also on the
2797 * reader_page, we let it move out.
2799 if (unlikely((cpu_buffer->commit_page !=
2800 cpu_buffer->tail_page) &&
2801 (cpu_buffer->commit_page ==
2802 cpu_buffer->reader_page))) {
2803 local_inc(&cpu_buffer->commit_overrun);
2809 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2813 rb_reset_tail(cpu_buffer, tail, info);
2815 /* Commit what we have for now. */
2816 rb_end_commit(cpu_buffer);
2817 /* rb_end_commit() decs committing */
2818 local_inc(&cpu_buffer->committing);
2820 /* fail and let the caller try again */
2821 return ERR_PTR(-EAGAIN);
2825 rb_reset_tail(cpu_buffer, tail, info);
2831 static struct ring_buffer_event *
2832 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2835 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2837 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2839 /* Not the first event on the page, or not delta? */
2840 if (abs || rb_event_index(event)) {
2841 event->time_delta = delta & TS_MASK;
2842 event->array[0] = delta >> TS_SHIFT;
2844 /* nope, just zero it */
2845 event->time_delta = 0;
2846 event->array[0] = 0;
2849 return skip_time_extend(event);
2852 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2853 static inline bool sched_clock_stable(void)
2860 rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2861 struct rb_event_info *info)
2865 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2866 (unsigned long long)info->delta,
2867 (unsigned long long)info->ts,
2868 (unsigned long long)info->before,
2869 (unsigned long long)info->after,
2870 (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2871 sched_clock_stable() ? "" :
2872 "If you just came from a suspend/resume,\n"
2873 "please switch to the trace global clock:\n"
2874 " echo global > /sys/kernel/tracing/trace_clock\n"
2875 "or add trace_clock=global to the kernel command line\n");
2878 static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2879 struct ring_buffer_event **event,
2880 struct rb_event_info *info,
2882 unsigned int *length)
2884 bool abs = info->add_timestamp &
2885 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2887 if (unlikely(info->delta > (1ULL << 59))) {
2889 * Some timers can use more than 59 bits, and when a timestamp
2890 * is added to the buffer, it will lose those bits.
2892 if (abs && (info->ts & TS_MSB)) {
2893 info->delta &= ABS_TS_MASK;
2895 /* did the clock go backwards */
2896 } else if (info->before == info->after && info->before > info->ts) {
2897 /* not interrupted */
2901 * This is possible with a recalibrating of the TSC.
2902 * Do not produce a call stack, but just report it.
2906 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2907 info->before, info->ts);
2910 rb_check_timestamp(cpu_buffer, info);
2914 *event = rb_add_time_stamp(*event, info->delta, abs);
2915 *length -= RB_LEN_TIME_EXTEND;
2920 * rb_update_event - update event type and data
2921 * @cpu_buffer: The per cpu buffer of the @event
2922 * @event: the event to update
2923 * @info: The info to update the @event with (contains length and delta)
2925 * Update the type and data fields of the @event. The length
2926 * is the actual size that is written to the ring buffer,
2927 * and with this, we can determine what to place into the
2931 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2932 struct ring_buffer_event *event,
2933 struct rb_event_info *info)
2935 unsigned length = info->length;
2936 u64 delta = info->delta;
2937 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2939 if (!WARN_ON_ONCE(nest >= MAX_NEST))
2940 cpu_buffer->event_stamp[nest] = info->ts;
2943 * If we need to add a timestamp, then we
2944 * add it to the start of the reserved space.
2946 if (unlikely(info->add_timestamp))
2947 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2949 event->time_delta = delta;
2950 length -= RB_EVNT_HDR_SIZE;
2951 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2952 event->type_len = 0;
2953 event->array[0] = length;
2955 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2958 static unsigned rb_calculate_event_length(unsigned length)
2960 struct ring_buffer_event event; /* Used only for sizeof array */
2962 /* zero length can cause confusions */
2966 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2967 length += sizeof(event.array[0]);
2969 length += RB_EVNT_HDR_SIZE;
2970 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2973 * In case the time delta is larger than the 27 bits for it
2974 * in the header, we need to add a timestamp. If another
2975 * event comes in when trying to discard this one to increase
2976 * the length, then the timestamp will be added in the allocated
2977 * space of this event. If length is bigger than the size needed
2978 * for the TIME_EXTEND, then padding has to be used. The events
2979 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2980 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2981 * As length is a multiple of 4, we only need to worry if it
2982 * is 12 (RB_LEN_TIME_EXTEND + 4).
2984 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2985 length += RB_ALIGNMENT;
2990 static u64 rb_time_delta(struct ring_buffer_event *event)
2992 switch (event->type_len) {
2993 case RINGBUF_TYPE_PADDING:
2996 case RINGBUF_TYPE_TIME_EXTEND:
2997 return rb_event_time_stamp(event);
2999 case RINGBUF_TYPE_TIME_STAMP:
3002 case RINGBUF_TYPE_DATA:
3003 return event->time_delta;
3010 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
3011 struct ring_buffer_event *event)
3013 unsigned long new_index, old_index;
3014 struct buffer_page *bpage;
3019 new_index = rb_event_index(event);
3020 old_index = new_index + rb_event_ts_length(event);
3021 addr = (unsigned long)event;
3024 bpage = READ_ONCE(cpu_buffer->tail_page);
3026 delta = rb_time_delta(event);
3028 if (!rb_time_read(&cpu_buffer->write_stamp, &write_stamp))
3031 /* Make sure the write stamp is read before testing the location */
3034 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
3035 unsigned long write_mask =
3036 local_read(&bpage->write) & ~RB_WRITE_MASK;
3037 unsigned long event_length = rb_event_length(event);
3040 * For the before_stamp to be different than the write_stamp
3041 * to make sure that the next event adds an absolute
3042 * value and does not rely on the saved write stamp, which
3043 * is now going to be bogus.
3045 rb_time_set(&cpu_buffer->before_stamp, 0);
3047 /* Something came in, can't discard */
3048 if (!rb_time_cmpxchg(&cpu_buffer->write_stamp,
3049 write_stamp, write_stamp - delta))
3053 * If an event were to come in now, it would see that the
3054 * write_stamp and the before_stamp are different, and assume
3055 * that this event just added itself before updating
3056 * the write stamp. The interrupting event will fix the
3057 * write stamp for us, and use the before stamp as its delta.
3061 * This is on the tail page. It is possible that
3062 * a write could come in and move the tail page
3063 * and write to the next page. That is fine
3064 * because we just shorten what is on this page.
3066 old_index += write_mask;
3067 new_index += write_mask;
3069 /* caution: old_index gets updated on cmpxchg failure */
3070 if (local_try_cmpxchg(&bpage->write, &old_index, new_index)) {
3071 /* update counters */
3072 local_sub(event_length, &cpu_buffer->entries_bytes);
3077 /* could not discard */
3081 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
3083 local_inc(&cpu_buffer->committing);
3084 local_inc(&cpu_buffer->commits);
3087 static __always_inline void
3088 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
3090 unsigned long max_count;
3093 * We only race with interrupts and NMIs on this CPU.
3094 * If we own the commit event, then we can commit
3095 * all others that interrupted us, since the interruptions
3096 * are in stack format (they finish before they come
3097 * back to us). This allows us to do a simple loop to
3098 * assign the commit to the tail.
3101 max_count = cpu_buffer->nr_pages * 100;
3103 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
3104 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
3106 if (RB_WARN_ON(cpu_buffer,
3107 rb_is_reader_page(cpu_buffer->tail_page)))
3110 * No need for a memory barrier here, as the update
3111 * of the tail_page did it for this page.
3113 local_set(&cpu_buffer->commit_page->page->commit,
3114 rb_page_write(cpu_buffer->commit_page));
3115 rb_inc_page(&cpu_buffer->commit_page);
3116 /* add barrier to keep gcc from optimizing too much */
3119 while (rb_commit_index(cpu_buffer) !=
3120 rb_page_write(cpu_buffer->commit_page)) {
3122 /* Make sure the readers see the content of what is committed. */
3124 local_set(&cpu_buffer->commit_page->page->commit,
3125 rb_page_write(cpu_buffer->commit_page));
3126 RB_WARN_ON(cpu_buffer,
3127 local_read(&cpu_buffer->commit_page->page->commit) &
3132 /* again, keep gcc from optimizing */
3136 * If an interrupt came in just after the first while loop
3137 * and pushed the tail page forward, we will be left with
3138 * a dangling commit that will never go forward.
3140 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3144 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3146 unsigned long commits;
3148 if (RB_WARN_ON(cpu_buffer,
3149 !local_read(&cpu_buffer->committing)))
3153 commits = local_read(&cpu_buffer->commits);
3154 /* synchronize with interrupts */
3156 if (local_read(&cpu_buffer->committing) == 1)
3157 rb_set_commit_to_write(cpu_buffer);
3159 local_dec(&cpu_buffer->committing);
3161 /* synchronize with interrupts */
3165 * Need to account for interrupts coming in between the
3166 * updating of the commit page and the clearing of the
3167 * committing counter.
3169 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3170 !local_read(&cpu_buffer->committing)) {
3171 local_inc(&cpu_buffer->committing);
3176 static inline void rb_event_discard(struct ring_buffer_event *event)
3178 if (extended_time(event))
3179 event = skip_time_extend(event);
3181 /* array[0] holds the actual length for the discarded event */
3182 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3183 event->type_len = RINGBUF_TYPE_PADDING;
3184 /* time delta must be non zero */
3185 if (!event->time_delta)
3186 event->time_delta = 1;
3189 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
3191 local_inc(&cpu_buffer->entries);
3192 rb_end_commit(cpu_buffer);
3195 static __always_inline void
3196 rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3198 if (buffer->irq_work.waiters_pending) {
3199 buffer->irq_work.waiters_pending = false;
3200 /* irq_work_queue() supplies it's own memory barriers */
3201 irq_work_queue(&buffer->irq_work.work);
3204 if (cpu_buffer->irq_work.waiters_pending) {
3205 cpu_buffer->irq_work.waiters_pending = false;
3206 /* irq_work_queue() supplies it's own memory barriers */
3207 irq_work_queue(&cpu_buffer->irq_work.work);
3210 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3213 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3216 if (!cpu_buffer->irq_work.full_waiters_pending)
3219 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3221 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3224 cpu_buffer->irq_work.wakeup_full = true;
3225 cpu_buffer->irq_work.full_waiters_pending = false;
3226 /* irq_work_queue() supplies it's own memory barriers */
3227 irq_work_queue(&cpu_buffer->irq_work.work);
3230 #ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3231 # define do_ring_buffer_record_recursion() \
3232 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3234 # define do_ring_buffer_record_recursion() do { } while (0)
3238 * The lock and unlock are done within a preempt disable section.
3239 * The current_context per_cpu variable can only be modified
3240 * by the current task between lock and unlock. But it can
3241 * be modified more than once via an interrupt. To pass this
3242 * information from the lock to the unlock without having to
3243 * access the 'in_interrupt()' functions again (which do show
3244 * a bit of overhead in something as critical as function tracing,
3245 * we use a bitmask trick.
3247 * bit 1 = NMI context
3248 * bit 2 = IRQ context
3249 * bit 3 = SoftIRQ context
3250 * bit 4 = normal context.
3252 * This works because this is the order of contexts that can
3253 * preempt other contexts. A SoftIRQ never preempts an IRQ
3256 * When the context is determined, the corresponding bit is
3257 * checked and set (if it was set, then a recursion of that context
3260 * On unlock, we need to clear this bit. To do so, just subtract
3261 * 1 from the current_context and AND it to itself.
3265 * 101 & 100 = 100 (clearing bit zero)
3268 * 1010 & 1001 = 1000 (clearing bit 1)
3270 * The least significant bit can be cleared this way, and it
3271 * just so happens that it is the same bit corresponding to
3272 * the current context.
3274 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3275 * is set when a recursion is detected at the current context, and if
3276 * the TRANSITION bit is already set, it will fail the recursion.
3277 * This is needed because there's a lag between the changing of
3278 * interrupt context and updating the preempt count. In this case,
3279 * a false positive will be found. To handle this, one extra recursion
3280 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3281 * bit is already set, then it is considered a recursion and the function
3282 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3284 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3285 * to be cleared. Even if it wasn't the context that set it. That is,
3286 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3287 * is called before preempt_count() is updated, since the check will
3288 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3289 * NMI then comes in, it will set the NMI bit, but when the NMI code
3290 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3291 * and leave the NMI bit set. But this is fine, because the interrupt
3292 * code that set the TRANSITION bit will then clear the NMI bit when it
3293 * calls trace_recursive_unlock(). If another NMI comes in, it will
3294 * set the TRANSITION bit and continue.
3296 * Note: The TRANSITION bit only handles a single transition between context.
3299 static __always_inline bool
3300 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3302 unsigned int val = cpu_buffer->current_context;
3303 int bit = interrupt_context_level();
3305 bit = RB_CTX_NORMAL - bit;
3307 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3309 * It is possible that this was called by transitioning
3310 * between interrupt context, and preempt_count() has not
3311 * been updated yet. In this case, use the TRANSITION bit.
3313 bit = RB_CTX_TRANSITION;
3314 if (val & (1 << (bit + cpu_buffer->nest))) {
3315 do_ring_buffer_record_recursion();
3320 val |= (1 << (bit + cpu_buffer->nest));
3321 cpu_buffer->current_context = val;
3326 static __always_inline void
3327 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3329 cpu_buffer->current_context &=
3330 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3333 /* The recursive locking above uses 5 bits */
3334 #define NESTED_BITS 5
3337 * ring_buffer_nest_start - Allow to trace while nested
3338 * @buffer: The ring buffer to modify
3340 * The ring buffer has a safety mechanism to prevent recursion.
3341 * But there may be a case where a trace needs to be done while
3342 * tracing something else. In this case, calling this function
3343 * will allow this function to nest within a currently active
3344 * ring_buffer_lock_reserve().
3346 * Call this function before calling another ring_buffer_lock_reserve() and
3347 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3349 void ring_buffer_nest_start(struct trace_buffer *buffer)
3351 struct ring_buffer_per_cpu *cpu_buffer;
3354 /* Enabled by ring_buffer_nest_end() */
3355 preempt_disable_notrace();
3356 cpu = raw_smp_processor_id();
3357 cpu_buffer = buffer->buffers[cpu];
3358 /* This is the shift value for the above recursive locking */
3359 cpu_buffer->nest += NESTED_BITS;
3363 * ring_buffer_nest_end - Allow to trace while nested
3364 * @buffer: The ring buffer to modify
3366 * Must be called after ring_buffer_nest_start() and after the
3367 * ring_buffer_unlock_commit().
3369 void ring_buffer_nest_end(struct trace_buffer *buffer)
3371 struct ring_buffer_per_cpu *cpu_buffer;
3374 /* disabled by ring_buffer_nest_start() */
3375 cpu = raw_smp_processor_id();
3376 cpu_buffer = buffer->buffers[cpu];
3377 /* This is the shift value for the above recursive locking */
3378 cpu_buffer->nest -= NESTED_BITS;
3379 preempt_enable_notrace();
3383 * ring_buffer_unlock_commit - commit a reserved
3384 * @buffer: The buffer to commit to
3386 * This commits the data to the ring buffer, and releases any locks held.
3388 * Must be paired with ring_buffer_lock_reserve.
3390 int ring_buffer_unlock_commit(struct trace_buffer *buffer)
3392 struct ring_buffer_per_cpu *cpu_buffer;
3393 int cpu = raw_smp_processor_id();
3395 cpu_buffer = buffer->buffers[cpu];
3397 rb_commit(cpu_buffer);
3399 rb_wakeups(buffer, cpu_buffer);
3401 trace_recursive_unlock(cpu_buffer);
3403 preempt_enable_notrace();
3407 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3409 /* Special value to validate all deltas on a page. */
3410 #define CHECK_FULL_PAGE 1L
3412 #ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3413 static void dump_buffer_page(struct buffer_data_page *bpage,
3414 struct rb_event_info *info,
3417 struct ring_buffer_event *event;
3421 ts = bpage->time_stamp;
3422 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
3424 for (e = 0; e < tail; e += rb_event_length(event)) {
3426 event = (struct ring_buffer_event *)(bpage->data + e);
3428 switch (event->type_len) {
3430 case RINGBUF_TYPE_TIME_EXTEND:
3431 delta = rb_event_time_stamp(event);
3433 pr_warn(" [%lld] delta:%lld TIME EXTEND\n", ts, delta);
3436 case RINGBUF_TYPE_TIME_STAMP:
3437 delta = rb_event_time_stamp(event);
3438 ts = rb_fix_abs_ts(delta, ts);
3439 pr_warn(" [%lld] absolute:%lld TIME STAMP\n", ts, delta);
3442 case RINGBUF_TYPE_PADDING:
3443 ts += event->time_delta;
3444 pr_warn(" [%lld] delta:%d PADDING\n", ts, event->time_delta);
3447 case RINGBUF_TYPE_DATA:
3448 ts += event->time_delta;
3449 pr_warn(" [%lld] delta:%d\n", ts, event->time_delta);
3458 static DEFINE_PER_CPU(atomic_t, checking);
3459 static atomic_t ts_dump;
3462 * Check if the current event time stamp matches the deltas on
3465 static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3466 struct rb_event_info *info,
3469 struct ring_buffer_event *event;
3470 struct buffer_data_page *bpage;
3475 bpage = info->tail_page->page;
3477 if (tail == CHECK_FULL_PAGE) {
3479 tail = local_read(&bpage->commit);
3480 } else if (info->add_timestamp &
3481 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3482 /* Ignore events with absolute time stamps */
3487 * Do not check the first event (skip possible extends too).
3488 * Also do not check if previous events have not been committed.
3490 if (tail <= 8 || tail > local_read(&bpage->commit))
3494 * If this interrupted another event,
3496 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3499 ts = bpage->time_stamp;
3501 for (e = 0; e < tail; e += rb_event_length(event)) {
3503 event = (struct ring_buffer_event *)(bpage->data + e);
3505 switch (event->type_len) {
3507 case RINGBUF_TYPE_TIME_EXTEND:
3508 delta = rb_event_time_stamp(event);
3512 case RINGBUF_TYPE_TIME_STAMP:
3513 delta = rb_event_time_stamp(event);
3514 ts = rb_fix_abs_ts(delta, ts);
3517 case RINGBUF_TYPE_PADDING:
3518 if (event->time_delta == 1)
3521 case RINGBUF_TYPE_DATA:
3522 ts += event->time_delta;
3526 RB_WARN_ON(cpu_buffer, 1);
3529 if ((full && ts > info->ts) ||
3530 (!full && ts + info->delta != info->ts)) {
3531 /* If another report is happening, ignore this one */
3532 if (atomic_inc_return(&ts_dump) != 1) {
3533 atomic_dec(&ts_dump);
3536 atomic_inc(&cpu_buffer->record_disabled);
3537 /* There's some cases in boot up that this can happen */
3538 WARN_ON_ONCE(system_state != SYSTEM_BOOTING);
3539 pr_warn("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s\n",
3541 ts + info->delta, info->ts, info->delta,
3542 info->before, info->after,
3543 full ? " (full)" : "");
3544 dump_buffer_page(bpage, info, tail);
3545 atomic_dec(&ts_dump);
3546 /* Do not re-enable checking */
3550 atomic_dec(this_cpu_ptr(&checking));
3553 static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3554 struct rb_event_info *info,
3558 #endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3560 static struct ring_buffer_event *
3561 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3562 struct rb_event_info *info)
3564 struct ring_buffer_event *event;
3565 struct buffer_page *tail_page;
3566 unsigned long tail, write, w;
3570 /* Don't let the compiler play games with cpu_buffer->tail_page */
3571 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3573 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3575 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3576 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3578 info->ts = rb_time_stamp(cpu_buffer->buffer);
3580 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3581 info->delta = info->ts;
3584 * If interrupting an event time update, we may need an
3585 * absolute timestamp.
3586 * Don't bother if this is the start of a new page (w == 0).
3589 /* Use the sub-buffer timestamp */
3591 } else if (unlikely(!a_ok || !b_ok || info->before != info->after)) {
3592 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3593 info->length += RB_LEN_TIME_EXTEND;
3595 info->delta = info->ts - info->after;
3596 if (unlikely(test_time_stamp(info->delta))) {
3597 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3598 info->length += RB_LEN_TIME_EXTEND;
3603 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3605 /*C*/ write = local_add_return(info->length, &tail_page->write);
3607 /* set write to only the index of the write */
3608 write &= RB_WRITE_MASK;
3610 tail = write - info->length;
3612 /* See if we shot pass the end of this buffer page */
3613 if (unlikely(write > BUF_PAGE_SIZE)) {
3614 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3615 return rb_move_tail(cpu_buffer, tail, info);
3618 if (likely(tail == w)) {
3619 /* Nothing interrupted us between A and C */
3620 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3622 * If something came in between C and D, the write stamp
3623 * may now not be in sync. But that's fine as the before_stamp
3624 * will be different and then next event will just be forced
3625 * to use an absolute timestamp.
3627 if (likely(!(info->add_timestamp &
3628 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3629 /* This did not interrupt any time update */
3630 info->delta = info->ts - info->after;
3632 /* Just use full timestamp for interrupting event */
3633 info->delta = info->ts;
3634 check_buffer(cpu_buffer, info, tail);
3637 /* SLOW PATH - Interrupted between A and C */
3638 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3639 /* Was interrupted before here, write_stamp must be valid */
3640 RB_WARN_ON(cpu_buffer, !a_ok);
3641 ts = rb_time_stamp(cpu_buffer->buffer);
3643 /*E*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3645 rb_time_cmpxchg(&cpu_buffer->write_stamp,
3647 /* Nothing came after this event between C and E */
3648 info->delta = ts - info->after;
3651 * Interrupted between C and E:
3652 * Lost the previous events time stamp. Just set the
3653 * delta to zero, and this will be the same time as
3654 * the event this event interrupted. And the events that
3655 * came after this will still be correct (as they would
3656 * have built their delta on the previous event.
3661 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3665 * If this is the first commit on the page, then it has the same
3666 * timestamp as the page itself.
3668 if (unlikely(!tail && !(info->add_timestamp &
3669 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3672 /* We reserved something on the buffer */
3674 event = __rb_page_index(tail_page, tail);
3675 rb_update_event(cpu_buffer, event, info);
3677 local_inc(&tail_page->entries);
3680 * If this is the first commit on the page, then update
3683 if (unlikely(!tail))
3684 tail_page->page->time_stamp = info->ts;
3686 /* account for these added bytes */
3687 local_add(info->length, &cpu_buffer->entries_bytes);
3692 static __always_inline struct ring_buffer_event *
3693 rb_reserve_next_event(struct trace_buffer *buffer,
3694 struct ring_buffer_per_cpu *cpu_buffer,
3695 unsigned long length)
3697 struct ring_buffer_event *event;
3698 struct rb_event_info info;
3702 rb_start_commit(cpu_buffer);
3703 /* The commit page can not change after this */
3705 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3707 * Due to the ability to swap a cpu buffer from a buffer
3708 * it is possible it was swapped before we committed.
3709 * (committing stops a swap). We check for it here and
3710 * if it happened, we have to fail the write.
3713 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3714 local_dec(&cpu_buffer->committing);
3715 local_dec(&cpu_buffer->commits);
3720 info.length = rb_calculate_event_length(length);
3722 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3723 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3724 info.length += RB_LEN_TIME_EXTEND;
3725 if (info.length > BUF_MAX_DATA_SIZE)
3728 add_ts_default = RB_ADD_STAMP_NONE;
3732 info.add_timestamp = add_ts_default;
3736 * We allow for interrupts to reenter here and do a trace.
3737 * If one does, it will cause this original code to loop
3738 * back here. Even with heavy interrupts happening, this
3739 * should only happen a few times in a row. If this happens
3740 * 1000 times in a row, there must be either an interrupt
3741 * storm or we have something buggy.
3744 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3747 event = __rb_reserve_next(cpu_buffer, &info);
3749 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3750 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3751 info.length -= RB_LEN_TIME_EXTEND;
3758 rb_end_commit(cpu_buffer);
3763 * ring_buffer_lock_reserve - reserve a part of the buffer
3764 * @buffer: the ring buffer to reserve from
3765 * @length: the length of the data to reserve (excluding event header)
3767 * Returns a reserved event on the ring buffer to copy directly to.
3768 * The user of this interface will need to get the body to write into
3769 * and can use the ring_buffer_event_data() interface.
3771 * The length is the length of the data needed, not the event length
3772 * which also includes the event header.
3774 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3775 * If NULL is returned, then nothing has been allocated or locked.
3777 struct ring_buffer_event *
3778 ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3780 struct ring_buffer_per_cpu *cpu_buffer;
3781 struct ring_buffer_event *event;
3784 /* If we are tracing schedule, we don't want to recurse */
3785 preempt_disable_notrace();
3787 if (unlikely(atomic_read(&buffer->record_disabled)))
3790 cpu = raw_smp_processor_id();
3792 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3795 cpu_buffer = buffer->buffers[cpu];
3797 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3800 if (unlikely(length > BUF_MAX_DATA_SIZE))
3803 if (unlikely(trace_recursive_lock(cpu_buffer)))
3806 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3813 trace_recursive_unlock(cpu_buffer);
3815 preempt_enable_notrace();
3818 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3821 * Decrement the entries to the page that an event is on.
3822 * The event does not even need to exist, only the pointer
3823 * to the page it is on. This may only be called before the commit
3827 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3828 struct ring_buffer_event *event)
3830 unsigned long addr = (unsigned long)event;
3831 struct buffer_page *bpage = cpu_buffer->commit_page;
3832 struct buffer_page *start;
3836 /* Do the likely case first */
3837 if (likely(bpage->page == (void *)addr)) {
3838 local_dec(&bpage->entries);
3843 * Because the commit page may be on the reader page we
3844 * start with the next page and check the end loop there.
3846 rb_inc_page(&bpage);
3849 if (bpage->page == (void *)addr) {
3850 local_dec(&bpage->entries);
3853 rb_inc_page(&bpage);
3854 } while (bpage != start);
3856 /* commit not part of this buffer?? */
3857 RB_WARN_ON(cpu_buffer, 1);
3861 * ring_buffer_discard_commit - discard an event that has not been committed
3862 * @buffer: the ring buffer
3863 * @event: non committed event to discard
3865 * Sometimes an event that is in the ring buffer needs to be ignored.
3866 * This function lets the user discard an event in the ring buffer
3867 * and then that event will not be read later.
3869 * This function only works if it is called before the item has been
3870 * committed. It will try to free the event from the ring buffer
3871 * if another event has not been added behind it.
3873 * If another event has been added behind it, it will set the event
3874 * up as discarded, and perform the commit.
3876 * If this function is called, do not call ring_buffer_unlock_commit on
3879 void ring_buffer_discard_commit(struct trace_buffer *buffer,
3880 struct ring_buffer_event *event)
3882 struct ring_buffer_per_cpu *cpu_buffer;
3885 /* The event is discarded regardless */
3886 rb_event_discard(event);
3888 cpu = smp_processor_id();
3889 cpu_buffer = buffer->buffers[cpu];
3892 * This must only be called if the event has not been
3893 * committed yet. Thus we can assume that preemption
3894 * is still disabled.
3896 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3898 rb_decrement_entry(cpu_buffer, event);
3899 if (rb_try_to_discard(cpu_buffer, event))
3903 rb_end_commit(cpu_buffer);
3905 trace_recursive_unlock(cpu_buffer);
3907 preempt_enable_notrace();
3910 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3913 * ring_buffer_write - write data to the buffer without reserving
3914 * @buffer: The ring buffer to write to.
3915 * @length: The length of the data being written (excluding the event header)
3916 * @data: The data to write to the buffer.
3918 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3919 * one function. If you already have the data to write to the buffer, it
3920 * may be easier to simply call this function.
3922 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3923 * and not the length of the event which would hold the header.
3925 int ring_buffer_write(struct trace_buffer *buffer,
3926 unsigned long length,
3929 struct ring_buffer_per_cpu *cpu_buffer;
3930 struct ring_buffer_event *event;
3935 preempt_disable_notrace();
3937 if (atomic_read(&buffer->record_disabled))
3940 cpu = raw_smp_processor_id();
3942 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3945 cpu_buffer = buffer->buffers[cpu];
3947 if (atomic_read(&cpu_buffer->record_disabled))
3950 if (length > BUF_MAX_DATA_SIZE)
3953 if (unlikely(trace_recursive_lock(cpu_buffer)))
3956 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3960 body = rb_event_data(event);
3962 memcpy(body, data, length);
3964 rb_commit(cpu_buffer);
3966 rb_wakeups(buffer, cpu_buffer);
3971 trace_recursive_unlock(cpu_buffer);
3974 preempt_enable_notrace();
3978 EXPORT_SYMBOL_GPL(ring_buffer_write);
3980 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3982 struct buffer_page *reader = cpu_buffer->reader_page;
3983 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3984 struct buffer_page *commit = cpu_buffer->commit_page;
3986 /* In case of error, head will be NULL */
3987 if (unlikely(!head))
3990 /* Reader should exhaust content in reader page */
3991 if (reader->read != rb_page_commit(reader))
3995 * If writers are committing on the reader page, knowing all
3996 * committed content has been read, the ring buffer is empty.
3998 if (commit == reader)
4002 * If writers are committing on a page other than reader page
4003 * and head page, there should always be content to read.
4009 * Writers are committing on the head page, we just need
4010 * to care about there're committed data, and the reader will
4011 * swap reader page with head page when it is to read data.
4013 return rb_page_commit(commit) == 0;
4017 * ring_buffer_record_disable - stop all writes into the buffer
4018 * @buffer: The ring buffer to stop writes to.
4020 * This prevents all writes to the buffer. Any attempt to write
4021 * to the buffer after this will fail and return NULL.
4023 * The caller should call synchronize_rcu() after this.
4025 void ring_buffer_record_disable(struct trace_buffer *buffer)
4027 atomic_inc(&buffer->record_disabled);
4029 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
4032 * ring_buffer_record_enable - enable writes to the buffer
4033 * @buffer: The ring buffer to enable writes
4035 * Note, multiple disables will need the same number of enables
4036 * to truly enable the writing (much like preempt_disable).
4038 void ring_buffer_record_enable(struct trace_buffer *buffer)
4040 atomic_dec(&buffer->record_disabled);
4042 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
4045 * ring_buffer_record_off - stop all writes into the buffer
4046 * @buffer: The ring buffer to stop writes to.
4048 * This prevents all writes to the buffer. Any attempt to write
4049 * to the buffer after this will fail and return NULL.
4051 * This is different than ring_buffer_record_disable() as
4052 * it works like an on/off switch, where as the disable() version
4053 * must be paired with a enable().
4055 void ring_buffer_record_off(struct trace_buffer *buffer)
4058 unsigned int new_rd;
4060 rd = atomic_read(&buffer->record_disabled);
4062 new_rd = rd | RB_BUFFER_OFF;
4063 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4065 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
4068 * ring_buffer_record_on - restart writes into the buffer
4069 * @buffer: The ring buffer to start writes to.
4071 * This enables all writes to the buffer that was disabled by
4072 * ring_buffer_record_off().
4074 * This is different than ring_buffer_record_enable() as
4075 * it works like an on/off switch, where as the enable() version
4076 * must be paired with a disable().
4078 void ring_buffer_record_on(struct trace_buffer *buffer)
4081 unsigned int new_rd;
4083 rd = atomic_read(&buffer->record_disabled);
4085 new_rd = rd & ~RB_BUFFER_OFF;
4086 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4088 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
4091 * ring_buffer_record_is_on - return true if the ring buffer can write
4092 * @buffer: The ring buffer to see if write is enabled
4094 * Returns true if the ring buffer is in a state that it accepts writes.
4096 bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4098 return !atomic_read(&buffer->record_disabled);
4102 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4103 * @buffer: The ring buffer to see if write is set enabled
4105 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4106 * Note that this does NOT mean it is in a writable state.
4108 * It may return true when the ring buffer has been disabled by
4109 * ring_buffer_record_disable(), as that is a temporary disabling of
4112 bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4114 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4118 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4119 * @buffer: The ring buffer to stop writes to.
4120 * @cpu: The CPU buffer to stop
4122 * This prevents all writes to the buffer. Any attempt to write
4123 * to the buffer after this will fail and return NULL.
4125 * The caller should call synchronize_rcu() after this.
4127 void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4129 struct ring_buffer_per_cpu *cpu_buffer;
4131 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4134 cpu_buffer = buffer->buffers[cpu];
4135 atomic_inc(&cpu_buffer->record_disabled);
4137 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4140 * ring_buffer_record_enable_cpu - enable writes to the buffer
4141 * @buffer: The ring buffer to enable writes
4142 * @cpu: The CPU to enable.
4144 * Note, multiple disables will need the same number of enables
4145 * to truly enable the writing (much like preempt_disable).
4147 void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4149 struct ring_buffer_per_cpu *cpu_buffer;
4151 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4154 cpu_buffer = buffer->buffers[cpu];
4155 atomic_dec(&cpu_buffer->record_disabled);
4157 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4160 * The total entries in the ring buffer is the running counter
4161 * of entries entered into the ring buffer, minus the sum of
4162 * the entries read from the ring buffer and the number of
4163 * entries that were overwritten.
4165 static inline unsigned long
4166 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4168 return local_read(&cpu_buffer->entries) -
4169 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4173 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4174 * @buffer: The ring buffer
4175 * @cpu: The per CPU buffer to read from.
4177 u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4179 unsigned long flags;
4180 struct ring_buffer_per_cpu *cpu_buffer;
4181 struct buffer_page *bpage;
4184 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4187 cpu_buffer = buffer->buffers[cpu];
4188 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4190 * if the tail is on reader_page, oldest time stamp is on the reader
4193 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4194 bpage = cpu_buffer->reader_page;
4196 bpage = rb_set_head_page(cpu_buffer);
4198 ret = bpage->page->time_stamp;
4199 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4203 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4206 * ring_buffer_bytes_cpu - get the number of bytes unconsumed in a cpu buffer
4207 * @buffer: The ring buffer
4208 * @cpu: The per CPU buffer to read from.
4210 unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4212 struct ring_buffer_per_cpu *cpu_buffer;
4215 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4218 cpu_buffer = buffer->buffers[cpu];
4219 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4223 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4226 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4227 * @buffer: The ring buffer
4228 * @cpu: The per CPU buffer to get the entries from.
4230 unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4232 struct ring_buffer_per_cpu *cpu_buffer;
4234 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4237 cpu_buffer = buffer->buffers[cpu];
4239 return rb_num_of_entries(cpu_buffer);
4241 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4244 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4245 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4246 * @buffer: The ring buffer
4247 * @cpu: The per CPU buffer to get the number of overruns from
4249 unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4251 struct ring_buffer_per_cpu *cpu_buffer;
4254 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4257 cpu_buffer = buffer->buffers[cpu];
4258 ret = local_read(&cpu_buffer->overrun);
4262 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4265 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4266 * commits failing due to the buffer wrapping around while there are uncommitted
4267 * events, such as during an interrupt storm.
4268 * @buffer: The ring buffer
4269 * @cpu: The per CPU buffer to get the number of overruns from
4272 ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4274 struct ring_buffer_per_cpu *cpu_buffer;
4277 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4280 cpu_buffer = buffer->buffers[cpu];
4281 ret = local_read(&cpu_buffer->commit_overrun);
4285 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4288 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4289 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4290 * @buffer: The ring buffer
4291 * @cpu: The per CPU buffer to get the number of overruns from
4294 ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4296 struct ring_buffer_per_cpu *cpu_buffer;
4299 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4302 cpu_buffer = buffer->buffers[cpu];
4303 ret = local_read(&cpu_buffer->dropped_events);
4307 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4310 * ring_buffer_read_events_cpu - get the number of events successfully read
4311 * @buffer: The ring buffer
4312 * @cpu: The per CPU buffer to get the number of events read
4315 ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4317 struct ring_buffer_per_cpu *cpu_buffer;
4319 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4322 cpu_buffer = buffer->buffers[cpu];
4323 return cpu_buffer->read;
4325 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4328 * ring_buffer_entries - get the number of entries in a buffer
4329 * @buffer: The ring buffer
4331 * Returns the total number of entries in the ring buffer
4334 unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4336 struct ring_buffer_per_cpu *cpu_buffer;
4337 unsigned long entries = 0;
4340 /* if you care about this being correct, lock the buffer */
4341 for_each_buffer_cpu(buffer, cpu) {
4342 cpu_buffer = buffer->buffers[cpu];
4343 entries += rb_num_of_entries(cpu_buffer);
4348 EXPORT_SYMBOL_GPL(ring_buffer_entries);
4351 * ring_buffer_overruns - get the number of overruns in buffer
4352 * @buffer: The ring buffer
4354 * Returns the total number of overruns in the ring buffer
4357 unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4359 struct ring_buffer_per_cpu *cpu_buffer;
4360 unsigned long overruns = 0;
4363 /* if you care about this being correct, lock the buffer */
4364 for_each_buffer_cpu(buffer, cpu) {
4365 cpu_buffer = buffer->buffers[cpu];
4366 overruns += local_read(&cpu_buffer->overrun);
4371 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4373 static void rb_iter_reset(struct ring_buffer_iter *iter)
4375 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4377 /* Iterator usage is expected to have record disabled */
4378 iter->head_page = cpu_buffer->reader_page;
4379 iter->head = cpu_buffer->reader_page->read;
4380 iter->next_event = iter->head;
4382 iter->cache_reader_page = iter->head_page;
4383 iter->cache_read = cpu_buffer->read;
4384 iter->cache_pages_removed = cpu_buffer->pages_removed;
4387 iter->read_stamp = cpu_buffer->read_stamp;
4388 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4390 iter->read_stamp = iter->head_page->page->time_stamp;
4391 iter->page_stamp = iter->read_stamp;
4396 * ring_buffer_iter_reset - reset an iterator
4397 * @iter: The iterator to reset
4399 * Resets the iterator, so that it will start from the beginning
4402 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4404 struct ring_buffer_per_cpu *cpu_buffer;
4405 unsigned long flags;
4410 cpu_buffer = iter->cpu_buffer;
4412 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4413 rb_iter_reset(iter);
4414 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4416 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4419 * ring_buffer_iter_empty - check if an iterator has no more to read
4420 * @iter: The iterator to check
4422 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4424 struct ring_buffer_per_cpu *cpu_buffer;
4425 struct buffer_page *reader;
4426 struct buffer_page *head_page;
4427 struct buffer_page *commit_page;
4428 struct buffer_page *curr_commit_page;
4433 cpu_buffer = iter->cpu_buffer;
4434 reader = cpu_buffer->reader_page;
4435 head_page = cpu_buffer->head_page;
4436 commit_page = cpu_buffer->commit_page;
4437 commit_ts = commit_page->page->time_stamp;
4440 * When the writer goes across pages, it issues a cmpxchg which
4441 * is a mb(), which will synchronize with the rmb here.
4442 * (see rb_tail_page_update())
4445 commit = rb_page_commit(commit_page);
4446 /* We want to make sure that the commit page doesn't change */
4449 /* Make sure commit page didn't change */
4450 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4451 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4453 /* If the commit page changed, then there's more data */
4454 if (curr_commit_page != commit_page ||
4455 curr_commit_ts != commit_ts)
4458 /* Still racy, as it may return a false positive, but that's OK */
4459 return ((iter->head_page == commit_page && iter->head >= commit) ||
4460 (iter->head_page == reader && commit_page == head_page &&
4461 head_page->read == commit &&
4462 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4464 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4467 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4468 struct ring_buffer_event *event)
4472 switch (event->type_len) {
4473 case RINGBUF_TYPE_PADDING:
4476 case RINGBUF_TYPE_TIME_EXTEND:
4477 delta = rb_event_time_stamp(event);
4478 cpu_buffer->read_stamp += delta;
4481 case RINGBUF_TYPE_TIME_STAMP:
4482 delta = rb_event_time_stamp(event);
4483 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
4484 cpu_buffer->read_stamp = delta;
4487 case RINGBUF_TYPE_DATA:
4488 cpu_buffer->read_stamp += event->time_delta;
4492 RB_WARN_ON(cpu_buffer, 1);
4497 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4498 struct ring_buffer_event *event)
4502 switch (event->type_len) {
4503 case RINGBUF_TYPE_PADDING:
4506 case RINGBUF_TYPE_TIME_EXTEND:
4507 delta = rb_event_time_stamp(event);
4508 iter->read_stamp += delta;
4511 case RINGBUF_TYPE_TIME_STAMP:
4512 delta = rb_event_time_stamp(event);
4513 delta = rb_fix_abs_ts(delta, iter->read_stamp);
4514 iter->read_stamp = delta;
4517 case RINGBUF_TYPE_DATA:
4518 iter->read_stamp += event->time_delta;
4522 RB_WARN_ON(iter->cpu_buffer, 1);
4526 static struct buffer_page *
4527 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4529 struct buffer_page *reader = NULL;
4530 unsigned long overwrite;
4531 unsigned long flags;
4535 local_irq_save(flags);
4536 arch_spin_lock(&cpu_buffer->lock);
4540 * This should normally only loop twice. But because the
4541 * start of the reader inserts an empty page, it causes
4542 * a case where we will loop three times. There should be no
4543 * reason to loop four times (that I know of).
4545 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4550 reader = cpu_buffer->reader_page;
4552 /* If there's more to read, return this page */
4553 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4556 /* Never should we have an index greater than the size */
4557 if (RB_WARN_ON(cpu_buffer,
4558 cpu_buffer->reader_page->read > rb_page_size(reader)))
4561 /* check if we caught up to the tail */
4563 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4566 /* Don't bother swapping if the ring buffer is empty */
4567 if (rb_num_of_entries(cpu_buffer) == 0)
4571 * Reset the reader page to size zero.
4573 local_set(&cpu_buffer->reader_page->write, 0);
4574 local_set(&cpu_buffer->reader_page->entries, 0);
4575 local_set(&cpu_buffer->reader_page->page->commit, 0);
4576 cpu_buffer->reader_page->real_end = 0;
4580 * Splice the empty reader page into the list around the head.
4582 reader = rb_set_head_page(cpu_buffer);
4585 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4586 cpu_buffer->reader_page->list.prev = reader->list.prev;
4589 * cpu_buffer->pages just needs to point to the buffer, it
4590 * has no specific buffer page to point to. Lets move it out
4591 * of our way so we don't accidentally swap it.
4593 cpu_buffer->pages = reader->list.prev;
4595 /* The reader page will be pointing to the new head */
4596 rb_set_list_to_head(&cpu_buffer->reader_page->list);
4599 * We want to make sure we read the overruns after we set up our
4600 * pointers to the next object. The writer side does a
4601 * cmpxchg to cross pages which acts as the mb on the writer
4602 * side. Note, the reader will constantly fail the swap
4603 * while the writer is updating the pointers, so this
4604 * guarantees that the overwrite recorded here is the one we
4605 * want to compare with the last_overrun.
4608 overwrite = local_read(&(cpu_buffer->overrun));
4611 * Here's the tricky part.
4613 * We need to move the pointer past the header page.
4614 * But we can only do that if a writer is not currently
4615 * moving it. The page before the header page has the
4616 * flag bit '1' set if it is pointing to the page we want.
4617 * but if the writer is in the process of moving it
4618 * than it will be '2' or already moved '0'.
4621 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4624 * If we did not convert it, then we must try again.
4630 * Yay! We succeeded in replacing the page.
4632 * Now make the new head point back to the reader page.
4634 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4635 rb_inc_page(&cpu_buffer->head_page);
4637 local_inc(&cpu_buffer->pages_read);
4639 /* Finally update the reader page to the new head */
4640 cpu_buffer->reader_page = reader;
4641 cpu_buffer->reader_page->read = 0;
4643 if (overwrite != cpu_buffer->last_overrun) {
4644 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4645 cpu_buffer->last_overrun = overwrite;
4651 /* Update the read_stamp on the first event */
4652 if (reader && reader->read == 0)
4653 cpu_buffer->read_stamp = reader->page->time_stamp;
4655 arch_spin_unlock(&cpu_buffer->lock);
4656 local_irq_restore(flags);
4659 * The writer has preempt disable, wait for it. But not forever
4660 * Although, 1 second is pretty much "forever"
4662 #define USECS_WAIT 1000000
4663 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4664 /* If the write is past the end of page, a writer is still updating it */
4665 if (likely(!reader || rb_page_write(reader) <= BUF_PAGE_SIZE))
4670 /* Get the latest version of the reader write value */
4674 /* The writer is not moving forward? Something is wrong */
4675 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4679 * Make sure we see any padding after the write update
4680 * (see rb_reset_tail()).
4682 * In addition, a writer may be writing on the reader page
4683 * if the page has not been fully filled, so the read barrier
4684 * is also needed to make sure we see the content of what is
4685 * committed by the writer (see rb_set_commit_to_write()).
4693 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4695 struct ring_buffer_event *event;
4696 struct buffer_page *reader;
4699 reader = rb_get_reader_page(cpu_buffer);
4701 /* This function should not be called when buffer is empty */
4702 if (RB_WARN_ON(cpu_buffer, !reader))
4705 event = rb_reader_event(cpu_buffer);
4707 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4710 rb_update_read_stamp(cpu_buffer, event);
4712 length = rb_event_length(event);
4713 cpu_buffer->reader_page->read += length;
4714 cpu_buffer->read_bytes += length;
4717 static void rb_advance_iter(struct ring_buffer_iter *iter)
4719 struct ring_buffer_per_cpu *cpu_buffer;
4721 cpu_buffer = iter->cpu_buffer;
4723 /* If head == next_event then we need to jump to the next event */
4724 if (iter->head == iter->next_event) {
4725 /* If the event gets overwritten again, there's nothing to do */
4726 if (rb_iter_head_event(iter) == NULL)
4730 iter->head = iter->next_event;
4733 * Check if we are at the end of the buffer.
4735 if (iter->next_event >= rb_page_size(iter->head_page)) {
4736 /* discarded commits can make the page empty */
4737 if (iter->head_page == cpu_buffer->commit_page)
4743 rb_update_iter_read_stamp(iter, iter->event);
4746 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4748 return cpu_buffer->lost_events;
4751 static struct ring_buffer_event *
4752 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4753 unsigned long *lost_events)
4755 struct ring_buffer_event *event;
4756 struct buffer_page *reader;
4763 * We repeat when a time extend is encountered.
4764 * Since the time extend is always attached to a data event,
4765 * we should never loop more than once.
4766 * (We never hit the following condition more than twice).
4768 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4771 reader = rb_get_reader_page(cpu_buffer);
4775 event = rb_reader_event(cpu_buffer);
4777 switch (event->type_len) {
4778 case RINGBUF_TYPE_PADDING:
4779 if (rb_null_event(event))
4780 RB_WARN_ON(cpu_buffer, 1);
4782 * Because the writer could be discarding every
4783 * event it creates (which would probably be bad)
4784 * if we were to go back to "again" then we may never
4785 * catch up, and will trigger the warn on, or lock
4786 * the box. Return the padding, and we will release
4787 * the current locks, and try again.
4791 case RINGBUF_TYPE_TIME_EXTEND:
4792 /* Internal data, OK to advance */
4793 rb_advance_reader(cpu_buffer);
4796 case RINGBUF_TYPE_TIME_STAMP:
4798 *ts = rb_event_time_stamp(event);
4799 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
4800 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4801 cpu_buffer->cpu, ts);
4803 /* Internal data, OK to advance */
4804 rb_advance_reader(cpu_buffer);
4807 case RINGBUF_TYPE_DATA:
4809 *ts = cpu_buffer->read_stamp + event->time_delta;
4810 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4811 cpu_buffer->cpu, ts);
4814 *lost_events = rb_lost_events(cpu_buffer);
4818 RB_WARN_ON(cpu_buffer, 1);
4823 EXPORT_SYMBOL_GPL(ring_buffer_peek);
4825 static struct ring_buffer_event *
4826 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4828 struct trace_buffer *buffer;
4829 struct ring_buffer_per_cpu *cpu_buffer;
4830 struct ring_buffer_event *event;
4836 cpu_buffer = iter->cpu_buffer;
4837 buffer = cpu_buffer->buffer;
4840 * Check if someone performed a consuming read to the buffer
4841 * or removed some pages from the buffer. In these cases,
4842 * iterator was invalidated and we need to reset it.
4844 if (unlikely(iter->cache_read != cpu_buffer->read ||
4845 iter->cache_reader_page != cpu_buffer->reader_page ||
4846 iter->cache_pages_removed != cpu_buffer->pages_removed))
4847 rb_iter_reset(iter);
4850 if (ring_buffer_iter_empty(iter))
4854 * As the writer can mess with what the iterator is trying
4855 * to read, just give up if we fail to get an event after
4856 * three tries. The iterator is not as reliable when reading
4857 * the ring buffer with an active write as the consumer is.
4858 * Do not warn if the three failures is reached.
4863 if (rb_per_cpu_empty(cpu_buffer))
4866 if (iter->head >= rb_page_size(iter->head_page)) {
4871 event = rb_iter_head_event(iter);
4875 switch (event->type_len) {
4876 case RINGBUF_TYPE_PADDING:
4877 if (rb_null_event(event)) {
4881 rb_advance_iter(iter);
4884 case RINGBUF_TYPE_TIME_EXTEND:
4885 /* Internal data, OK to advance */
4886 rb_advance_iter(iter);
4889 case RINGBUF_TYPE_TIME_STAMP:
4891 *ts = rb_event_time_stamp(event);
4892 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
4893 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4894 cpu_buffer->cpu, ts);
4896 /* Internal data, OK to advance */
4897 rb_advance_iter(iter);
4900 case RINGBUF_TYPE_DATA:
4902 *ts = iter->read_stamp + event->time_delta;
4903 ring_buffer_normalize_time_stamp(buffer,
4904 cpu_buffer->cpu, ts);
4909 RB_WARN_ON(cpu_buffer, 1);
4914 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4916 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4918 if (likely(!in_nmi())) {
4919 raw_spin_lock(&cpu_buffer->reader_lock);
4924 * If an NMI die dumps out the content of the ring buffer
4925 * trylock must be used to prevent a deadlock if the NMI
4926 * preempted a task that holds the ring buffer locks. If
4927 * we get the lock then all is fine, if not, then continue
4928 * to do the read, but this can corrupt the ring buffer,
4929 * so it must be permanently disabled from future writes.
4930 * Reading from NMI is a oneshot deal.
4932 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4935 /* Continue without locking, but disable the ring buffer */
4936 atomic_inc(&cpu_buffer->record_disabled);
4941 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4944 raw_spin_unlock(&cpu_buffer->reader_lock);
4948 * ring_buffer_peek - peek at the next event to be read
4949 * @buffer: The ring buffer to read
4950 * @cpu: The cpu to peak at
4951 * @ts: The timestamp counter of this event.
4952 * @lost_events: a variable to store if events were lost (may be NULL)
4954 * This will return the event that will be read next, but does
4955 * not consume the data.
4957 struct ring_buffer_event *
4958 ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4959 unsigned long *lost_events)
4961 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4962 struct ring_buffer_event *event;
4963 unsigned long flags;
4966 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4970 local_irq_save(flags);
4971 dolock = rb_reader_lock(cpu_buffer);
4972 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4973 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4974 rb_advance_reader(cpu_buffer);
4975 rb_reader_unlock(cpu_buffer, dolock);
4976 local_irq_restore(flags);
4978 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4984 /** ring_buffer_iter_dropped - report if there are dropped events
4985 * @iter: The ring buffer iterator
4987 * Returns true if there was dropped events since the last peek.
4989 bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4991 bool ret = iter->missed_events != 0;
4993 iter->missed_events = 0;
4996 EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
4999 * ring_buffer_iter_peek - peek at the next event to be read
5000 * @iter: The ring buffer iterator
5001 * @ts: The timestamp counter of this event.
5003 * This will return the event that will be read next, but does
5004 * not increment the iterator.
5006 struct ring_buffer_event *
5007 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
5009 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5010 struct ring_buffer_event *event;
5011 unsigned long flags;
5014 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5015 event = rb_iter_peek(iter, ts);
5016 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5018 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5025 * ring_buffer_consume - return an event and consume it
5026 * @buffer: The ring buffer to get the next event from
5027 * @cpu: the cpu to read the buffer from
5028 * @ts: a variable to store the timestamp (may be NULL)
5029 * @lost_events: a variable to store if events were lost (may be NULL)
5031 * Returns the next event in the ring buffer, and that event is consumed.
5032 * Meaning, that sequential reads will keep returning a different event,
5033 * and eventually empty the ring buffer if the producer is slower.
5035 struct ring_buffer_event *
5036 ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
5037 unsigned long *lost_events)
5039 struct ring_buffer_per_cpu *cpu_buffer;
5040 struct ring_buffer_event *event = NULL;
5041 unsigned long flags;
5045 /* might be called in atomic */
5048 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5051 cpu_buffer = buffer->buffers[cpu];
5052 local_irq_save(flags);
5053 dolock = rb_reader_lock(cpu_buffer);
5055 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5057 cpu_buffer->lost_events = 0;
5058 rb_advance_reader(cpu_buffer);
5061 rb_reader_unlock(cpu_buffer, dolock);
5062 local_irq_restore(flags);
5067 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5072 EXPORT_SYMBOL_GPL(ring_buffer_consume);
5075 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
5076 * @buffer: The ring buffer to read from
5077 * @cpu: The cpu buffer to iterate over
5078 * @flags: gfp flags to use for memory allocation
5080 * This performs the initial preparations necessary to iterate
5081 * through the buffer. Memory is allocated, buffer recording
5082 * is disabled, and the iterator pointer is returned to the caller.
5084 * Disabling buffer recording prevents the reading from being
5085 * corrupted. This is not a consuming read, so a producer is not
5088 * After a sequence of ring_buffer_read_prepare calls, the user is
5089 * expected to make at least one call to ring_buffer_read_prepare_sync.
5090 * Afterwards, ring_buffer_read_start is invoked to get things going
5093 * This overall must be paired with ring_buffer_read_finish.
5095 struct ring_buffer_iter *
5096 ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5098 struct ring_buffer_per_cpu *cpu_buffer;
5099 struct ring_buffer_iter *iter;
5101 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5104 iter = kzalloc(sizeof(*iter), flags);
5108 /* Holds the entire event: data and meta data */
5109 iter->event = kmalloc(BUF_PAGE_SIZE, flags);
5115 cpu_buffer = buffer->buffers[cpu];
5117 iter->cpu_buffer = cpu_buffer;
5119 atomic_inc(&cpu_buffer->resize_disabled);
5123 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5126 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5128 * All previously invoked ring_buffer_read_prepare calls to prepare
5129 * iterators will be synchronized. Afterwards, read_buffer_read_start
5130 * calls on those iterators are allowed.
5133 ring_buffer_read_prepare_sync(void)
5137 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5140 * ring_buffer_read_start - start a non consuming read of the buffer
5141 * @iter: The iterator returned by ring_buffer_read_prepare
5143 * This finalizes the startup of an iteration through the buffer.
5144 * The iterator comes from a call to ring_buffer_read_prepare and
5145 * an intervening ring_buffer_read_prepare_sync must have been
5148 * Must be paired with ring_buffer_read_finish.
5151 ring_buffer_read_start(struct ring_buffer_iter *iter)
5153 struct ring_buffer_per_cpu *cpu_buffer;
5154 unsigned long flags;
5159 cpu_buffer = iter->cpu_buffer;
5161 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5162 arch_spin_lock(&cpu_buffer->lock);
5163 rb_iter_reset(iter);
5164 arch_spin_unlock(&cpu_buffer->lock);
5165 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5167 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5170 * ring_buffer_read_finish - finish reading the iterator of the buffer
5171 * @iter: The iterator retrieved by ring_buffer_start
5173 * This re-enables the recording to the buffer, and frees the
5177 ring_buffer_read_finish(struct ring_buffer_iter *iter)
5179 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5180 unsigned long flags;
5183 * Ring buffer is disabled from recording, here's a good place
5184 * to check the integrity of the ring buffer.
5185 * Must prevent readers from trying to read, as the check
5186 * clears the HEAD page and readers require it.
5188 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5189 rb_check_pages(cpu_buffer);
5190 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5192 atomic_dec(&cpu_buffer->resize_disabled);
5196 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5199 * ring_buffer_iter_advance - advance the iterator to the next location
5200 * @iter: The ring buffer iterator
5202 * Move the location of the iterator such that the next read will
5203 * be the next location of the iterator.
5205 void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5207 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5208 unsigned long flags;
5210 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5212 rb_advance_iter(iter);
5214 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5216 EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5219 * ring_buffer_size - return the size of the ring buffer (in bytes)
5220 * @buffer: The ring buffer.
5221 * @cpu: The CPU to get ring buffer size from.
5223 unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5226 * Earlier, this method returned
5227 * BUF_PAGE_SIZE * buffer->nr_pages
5228 * Since the nr_pages field is now removed, we have converted this to
5229 * return the per cpu buffer value.
5231 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5234 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
5236 EXPORT_SYMBOL_GPL(ring_buffer_size);
5238 static void rb_clear_buffer_page(struct buffer_page *page)
5240 local_set(&page->write, 0);
5241 local_set(&page->entries, 0);
5242 rb_init_page(page->page);
5247 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5249 struct buffer_page *page;
5251 rb_head_page_deactivate(cpu_buffer);
5253 cpu_buffer->head_page
5254 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5255 rb_clear_buffer_page(cpu_buffer->head_page);
5256 list_for_each_entry(page, cpu_buffer->pages, list) {
5257 rb_clear_buffer_page(page);
5260 cpu_buffer->tail_page = cpu_buffer->head_page;
5261 cpu_buffer->commit_page = cpu_buffer->head_page;
5263 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5264 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5265 rb_clear_buffer_page(cpu_buffer->reader_page);
5267 local_set(&cpu_buffer->entries_bytes, 0);
5268 local_set(&cpu_buffer->overrun, 0);
5269 local_set(&cpu_buffer->commit_overrun, 0);
5270 local_set(&cpu_buffer->dropped_events, 0);
5271 local_set(&cpu_buffer->entries, 0);
5272 local_set(&cpu_buffer->committing, 0);
5273 local_set(&cpu_buffer->commits, 0);
5274 local_set(&cpu_buffer->pages_touched, 0);
5275 local_set(&cpu_buffer->pages_lost, 0);
5276 local_set(&cpu_buffer->pages_read, 0);
5277 cpu_buffer->last_pages_touch = 0;
5278 cpu_buffer->shortest_full = 0;
5279 cpu_buffer->read = 0;
5280 cpu_buffer->read_bytes = 0;
5282 rb_time_set(&cpu_buffer->write_stamp, 0);
5283 rb_time_set(&cpu_buffer->before_stamp, 0);
5285 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5287 cpu_buffer->lost_events = 0;
5288 cpu_buffer->last_overrun = 0;
5290 rb_head_page_activate(cpu_buffer);
5291 cpu_buffer->pages_removed = 0;
5294 /* Must have disabled the cpu buffer then done a synchronize_rcu */
5295 static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5297 unsigned long flags;
5299 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5301 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5304 arch_spin_lock(&cpu_buffer->lock);
5306 rb_reset_cpu(cpu_buffer);
5308 arch_spin_unlock(&cpu_buffer->lock);
5311 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5315 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5316 * @buffer: The ring buffer to reset a per cpu buffer of
5317 * @cpu: The CPU buffer to be reset
5319 void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5321 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5323 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5326 /* prevent another thread from changing buffer sizes */
5327 mutex_lock(&buffer->mutex);
5329 atomic_inc(&cpu_buffer->resize_disabled);
5330 atomic_inc(&cpu_buffer->record_disabled);
5332 /* Make sure all commits have finished */
5335 reset_disabled_cpu_buffer(cpu_buffer);
5337 atomic_dec(&cpu_buffer->record_disabled);
5338 atomic_dec(&cpu_buffer->resize_disabled);
5340 mutex_unlock(&buffer->mutex);
5342 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5344 /* Flag to ensure proper resetting of atomic variables */
5345 #define RESET_BIT (1 << 30)
5348 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
5349 * @buffer: The ring buffer to reset a per cpu buffer of
5351 void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5353 struct ring_buffer_per_cpu *cpu_buffer;
5356 /* prevent another thread from changing buffer sizes */
5357 mutex_lock(&buffer->mutex);
5359 for_each_online_buffer_cpu(buffer, cpu) {
5360 cpu_buffer = buffer->buffers[cpu];
5362 atomic_add(RESET_BIT, &cpu_buffer->resize_disabled);
5363 atomic_inc(&cpu_buffer->record_disabled);
5366 /* Make sure all commits have finished */
5369 for_each_buffer_cpu(buffer, cpu) {
5370 cpu_buffer = buffer->buffers[cpu];
5373 * If a CPU came online during the synchronize_rcu(), then
5376 if (!(atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT))
5379 reset_disabled_cpu_buffer(cpu_buffer);
5381 atomic_dec(&cpu_buffer->record_disabled);
5382 atomic_sub(RESET_BIT, &cpu_buffer->resize_disabled);
5385 mutex_unlock(&buffer->mutex);
5389 * ring_buffer_reset - reset a ring buffer
5390 * @buffer: The ring buffer to reset all cpu buffers
5392 void ring_buffer_reset(struct trace_buffer *buffer)
5394 struct ring_buffer_per_cpu *cpu_buffer;
5397 /* prevent another thread from changing buffer sizes */
5398 mutex_lock(&buffer->mutex);
5400 for_each_buffer_cpu(buffer, cpu) {
5401 cpu_buffer = buffer->buffers[cpu];
5403 atomic_inc(&cpu_buffer->resize_disabled);
5404 atomic_inc(&cpu_buffer->record_disabled);
5407 /* Make sure all commits have finished */
5410 for_each_buffer_cpu(buffer, cpu) {
5411 cpu_buffer = buffer->buffers[cpu];
5413 reset_disabled_cpu_buffer(cpu_buffer);
5415 atomic_dec(&cpu_buffer->record_disabled);
5416 atomic_dec(&cpu_buffer->resize_disabled);
5419 mutex_unlock(&buffer->mutex);
5421 EXPORT_SYMBOL_GPL(ring_buffer_reset);
5424 * ring_buffer_empty - is the ring buffer empty?
5425 * @buffer: The ring buffer to test
5427 bool ring_buffer_empty(struct trace_buffer *buffer)
5429 struct ring_buffer_per_cpu *cpu_buffer;
5430 unsigned long flags;
5435 /* yes this is racy, but if you don't like the race, lock the buffer */
5436 for_each_buffer_cpu(buffer, cpu) {
5437 cpu_buffer = buffer->buffers[cpu];
5438 local_irq_save(flags);
5439 dolock = rb_reader_lock(cpu_buffer);
5440 ret = rb_per_cpu_empty(cpu_buffer);
5441 rb_reader_unlock(cpu_buffer, dolock);
5442 local_irq_restore(flags);
5450 EXPORT_SYMBOL_GPL(ring_buffer_empty);
5453 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5454 * @buffer: The ring buffer
5455 * @cpu: The CPU buffer to test
5457 bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5459 struct ring_buffer_per_cpu *cpu_buffer;
5460 unsigned long flags;
5464 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5467 cpu_buffer = buffer->buffers[cpu];
5468 local_irq_save(flags);
5469 dolock = rb_reader_lock(cpu_buffer);
5470 ret = rb_per_cpu_empty(cpu_buffer);
5471 rb_reader_unlock(cpu_buffer, dolock);
5472 local_irq_restore(flags);
5476 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5478 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5480 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5481 * @buffer_a: One buffer to swap with
5482 * @buffer_b: The other buffer to swap with
5483 * @cpu: the CPU of the buffers to swap
5485 * This function is useful for tracers that want to take a "snapshot"
5486 * of a CPU buffer and has another back up buffer lying around.
5487 * it is expected that the tracer handles the cpu buffer not being
5488 * used at the moment.
5490 int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5491 struct trace_buffer *buffer_b, int cpu)
5493 struct ring_buffer_per_cpu *cpu_buffer_a;
5494 struct ring_buffer_per_cpu *cpu_buffer_b;
5497 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5498 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5501 cpu_buffer_a = buffer_a->buffers[cpu];
5502 cpu_buffer_b = buffer_b->buffers[cpu];
5504 /* At least make sure the two buffers are somewhat the same */
5505 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5510 if (atomic_read(&buffer_a->record_disabled))
5513 if (atomic_read(&buffer_b->record_disabled))
5516 if (atomic_read(&cpu_buffer_a->record_disabled))
5519 if (atomic_read(&cpu_buffer_b->record_disabled))
5523 * We can't do a synchronize_rcu here because this
5524 * function can be called in atomic context.
5525 * Normally this will be called from the same CPU as cpu.
5526 * If not it's up to the caller to protect this.
5528 atomic_inc(&cpu_buffer_a->record_disabled);
5529 atomic_inc(&cpu_buffer_b->record_disabled);
5532 if (local_read(&cpu_buffer_a->committing))
5534 if (local_read(&cpu_buffer_b->committing))
5538 * When resize is in progress, we cannot swap it because
5539 * it will mess the state of the cpu buffer.
5541 if (atomic_read(&buffer_a->resizing))
5543 if (atomic_read(&buffer_b->resizing))
5546 buffer_a->buffers[cpu] = cpu_buffer_b;
5547 buffer_b->buffers[cpu] = cpu_buffer_a;
5549 cpu_buffer_b->buffer = buffer_a;
5550 cpu_buffer_a->buffer = buffer_b;
5555 atomic_dec(&cpu_buffer_a->record_disabled);
5556 atomic_dec(&cpu_buffer_b->record_disabled);
5560 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5561 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5564 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5565 * @buffer: the buffer to allocate for.
5566 * @cpu: the cpu buffer to allocate.
5568 * This function is used in conjunction with ring_buffer_read_page.
5569 * When reading a full page from the ring buffer, these functions
5570 * can be used to speed up the process. The calling function should
5571 * allocate a few pages first with this function. Then when it
5572 * needs to get pages from the ring buffer, it passes the result
5573 * of this function into ring_buffer_read_page, which will swap
5574 * the page that was allocated, with the read page of the buffer.
5577 * The page allocated, or ERR_PTR
5579 void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5581 struct ring_buffer_per_cpu *cpu_buffer;
5582 struct buffer_data_page *bpage = NULL;
5583 unsigned long flags;
5586 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5587 return ERR_PTR(-ENODEV);
5589 cpu_buffer = buffer->buffers[cpu];
5590 local_irq_save(flags);
5591 arch_spin_lock(&cpu_buffer->lock);
5593 if (cpu_buffer->free_page) {
5594 bpage = cpu_buffer->free_page;
5595 cpu_buffer->free_page = NULL;
5598 arch_spin_unlock(&cpu_buffer->lock);
5599 local_irq_restore(flags);
5604 page = alloc_pages_node(cpu_to_node(cpu),
5605 GFP_KERNEL | __GFP_NORETRY, 0);
5607 return ERR_PTR(-ENOMEM);
5609 bpage = page_address(page);
5612 rb_init_page(bpage);
5616 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5619 * ring_buffer_free_read_page - free an allocated read page
5620 * @buffer: the buffer the page was allocate for
5621 * @cpu: the cpu buffer the page came from
5622 * @data: the page to free
5624 * Free a page allocated from ring_buffer_alloc_read_page.
5626 void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5628 struct ring_buffer_per_cpu *cpu_buffer;
5629 struct buffer_data_page *bpage = data;
5630 struct page *page = virt_to_page(bpage);
5631 unsigned long flags;
5633 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
5636 cpu_buffer = buffer->buffers[cpu];
5638 /* If the page is still in use someplace else, we can't reuse it */
5639 if (page_ref_count(page) > 1)
5642 local_irq_save(flags);
5643 arch_spin_lock(&cpu_buffer->lock);
5645 if (!cpu_buffer->free_page) {
5646 cpu_buffer->free_page = bpage;
5650 arch_spin_unlock(&cpu_buffer->lock);
5651 local_irq_restore(flags);
5654 free_page((unsigned long)bpage);
5656 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5659 * ring_buffer_read_page - extract a page from the ring buffer
5660 * @buffer: buffer to extract from
5661 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5662 * @len: amount to extract
5663 * @cpu: the cpu of the buffer to extract
5664 * @full: should the extraction only happen when the page is full.
5666 * This function will pull out a page from the ring buffer and consume it.
5667 * @data_page must be the address of the variable that was returned
5668 * from ring_buffer_alloc_read_page. This is because the page might be used
5669 * to swap with a page in the ring buffer.
5672 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5673 * if (IS_ERR(rpage))
5674 * return PTR_ERR(rpage);
5675 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5677 * process_page(rpage, ret);
5679 * When @full is set, the function will not return true unless
5680 * the writer is off the reader page.
5682 * Note: it is up to the calling functions to handle sleeps and wakeups.
5683 * The ring buffer can be used anywhere in the kernel and can not
5684 * blindly call wake_up. The layer that uses the ring buffer must be
5685 * responsible for that.
5688 * >=0 if data has been transferred, returns the offset of consumed data.
5689 * <0 if no data has been transferred.
5691 int ring_buffer_read_page(struct trace_buffer *buffer,
5692 void **data_page, size_t len, int cpu, int full)
5694 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5695 struct ring_buffer_event *event;
5696 struct buffer_data_page *bpage;
5697 struct buffer_page *reader;
5698 unsigned long missed_events;
5699 unsigned long flags;
5700 unsigned int commit;
5705 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5709 * If len is not big enough to hold the page header, then
5710 * we can not copy anything.
5712 if (len <= BUF_PAGE_HDR_SIZE)
5715 len -= BUF_PAGE_HDR_SIZE;
5724 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5726 reader = rb_get_reader_page(cpu_buffer);
5730 event = rb_reader_event(cpu_buffer);
5732 read = reader->read;
5733 commit = rb_page_commit(reader);
5735 /* Check if any events were dropped */
5736 missed_events = cpu_buffer->lost_events;
5739 * If this page has been partially read or
5740 * if len is not big enough to read the rest of the page or
5741 * a writer is still on the page, then
5742 * we must copy the data from the page to the buffer.
5743 * Otherwise, we can simply swap the page with the one passed in.
5745 if (read || (len < (commit - read)) ||
5746 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5747 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5748 unsigned int rpos = read;
5749 unsigned int pos = 0;
5753 * If a full page is expected, this can still be returned
5754 * if there's been a previous partial read and the
5755 * rest of the page can be read and the commit page is off
5759 (!read || (len < (commit - read)) ||
5760 cpu_buffer->reader_page == cpu_buffer->commit_page))
5763 if (len > (commit - read))
5764 len = (commit - read);
5766 /* Always keep the time extend and data together */
5767 size = rb_event_ts_length(event);
5772 /* save the current timestamp, since the user will need it */
5773 save_timestamp = cpu_buffer->read_stamp;
5775 /* Need to copy one event at a time */
5777 /* We need the size of one event, because
5778 * rb_advance_reader only advances by one event,
5779 * whereas rb_event_ts_length may include the size of
5780 * one or two events.
5781 * We have already ensured there's enough space if this
5782 * is a time extend. */
5783 size = rb_event_length(event);
5784 memcpy(bpage->data + pos, rpage->data + rpos, size);
5788 rb_advance_reader(cpu_buffer);
5789 rpos = reader->read;
5795 event = rb_reader_event(cpu_buffer);
5796 /* Always keep the time extend and data together */
5797 size = rb_event_ts_length(event);
5798 } while (len >= size);
5801 local_set(&bpage->commit, pos);
5802 bpage->time_stamp = save_timestamp;
5804 /* we copied everything to the beginning */
5807 /* update the entry counter */
5808 cpu_buffer->read += rb_page_entries(reader);
5809 cpu_buffer->read_bytes += rb_page_commit(reader);
5811 /* swap the pages */
5812 rb_init_page(bpage);
5813 bpage = reader->page;
5814 reader->page = *data_page;
5815 local_set(&reader->write, 0);
5816 local_set(&reader->entries, 0);
5821 * Use the real_end for the data size,
5822 * This gives us a chance to store the lost events
5825 if (reader->real_end)
5826 local_set(&bpage->commit, reader->real_end);
5830 cpu_buffer->lost_events = 0;
5832 commit = local_read(&bpage->commit);
5834 * Set a flag in the commit field if we lost events
5836 if (missed_events) {
5837 /* If there is room at the end of the page to save the
5838 * missed events, then record it there.
5840 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5841 memcpy(&bpage->data[commit], &missed_events,
5842 sizeof(missed_events));
5843 local_add(RB_MISSED_STORED, &bpage->commit);
5844 commit += sizeof(missed_events);
5846 local_add(RB_MISSED_EVENTS, &bpage->commit);
5850 * This page may be off to user land. Zero it out here.
5852 if (commit < BUF_PAGE_SIZE)
5853 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5856 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5861 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5864 * We only allocate new buffers, never free them if the CPU goes down.
5865 * If we were to free the buffer, then the user would lose any trace that was in
5868 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5870 struct trace_buffer *buffer;
5873 unsigned long nr_pages;
5875 buffer = container_of(node, struct trace_buffer, node);
5876 if (cpumask_test_cpu(cpu, buffer->cpumask))
5881 /* check if all cpu sizes are same */
5882 for_each_buffer_cpu(buffer, cpu_i) {
5883 /* fill in the size from first enabled cpu */
5885 nr_pages = buffer->buffers[cpu_i]->nr_pages;
5886 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5891 /* allocate minimum pages, user can later expand it */
5894 buffer->buffers[cpu] =
5895 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5896 if (!buffer->buffers[cpu]) {
5897 WARN(1, "failed to allocate ring buffer on CPU %u\n",
5902 cpumask_set_cpu(cpu, buffer->cpumask);
5906 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5908 * This is a basic integrity check of the ring buffer.
5909 * Late in the boot cycle this test will run when configured in.
5910 * It will kick off a thread per CPU that will go into a loop
5911 * writing to the per cpu ring buffer various sizes of data.
5912 * Some of the data will be large items, some small.
5914 * Another thread is created that goes into a spin, sending out
5915 * IPIs to the other CPUs to also write into the ring buffer.
5916 * this is to test the nesting ability of the buffer.
5918 * Basic stats are recorded and reported. If something in the
5919 * ring buffer should happen that's not expected, a big warning
5920 * is displayed and all ring buffers are disabled.
5922 static struct task_struct *rb_threads[NR_CPUS] __initdata;
5924 struct rb_test_data {
5925 struct trace_buffer *buffer;
5926 unsigned long events;
5927 unsigned long bytes_written;
5928 unsigned long bytes_alloc;
5929 unsigned long bytes_dropped;
5930 unsigned long events_nested;
5931 unsigned long bytes_written_nested;
5932 unsigned long bytes_alloc_nested;
5933 unsigned long bytes_dropped_nested;
5934 int min_size_nested;
5935 int max_size_nested;
5942 static struct rb_test_data rb_data[NR_CPUS] __initdata;
5945 #define RB_TEST_BUFFER_SIZE 1048576
5947 static char rb_string[] __initdata =
5948 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5949 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5950 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5952 static bool rb_test_started __initdata;
5959 static __init int rb_write_something(struct rb_test_data *data, bool nested)
5961 struct ring_buffer_event *event;
5962 struct rb_item *item;
5969 /* Have nested writes different that what is written */
5970 cnt = data->cnt + (nested ? 27 : 0);
5972 /* Multiply cnt by ~e, to make some unique increment */
5973 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5975 len = size + sizeof(struct rb_item);
5977 started = rb_test_started;
5978 /* read rb_test_started before checking buffer enabled */
5981 event = ring_buffer_lock_reserve(data->buffer, len);
5983 /* Ignore dropped events before test starts. */
5986 data->bytes_dropped += len;
5988 data->bytes_dropped_nested += len;
5993 event_len = ring_buffer_event_length(event);
5995 if (RB_WARN_ON(data->buffer, event_len < len))
5998 item = ring_buffer_event_data(event);
6000 memcpy(item->str, rb_string, size);
6003 data->bytes_alloc_nested += event_len;
6004 data->bytes_written_nested += len;
6005 data->events_nested++;
6006 if (!data->min_size_nested || len < data->min_size_nested)
6007 data->min_size_nested = len;
6008 if (len > data->max_size_nested)
6009 data->max_size_nested = len;
6011 data->bytes_alloc += event_len;
6012 data->bytes_written += len;
6014 if (!data->min_size || len < data->min_size)
6015 data->max_size = len;
6016 if (len > data->max_size)
6017 data->max_size = len;
6021 ring_buffer_unlock_commit(data->buffer);
6026 static __init int rb_test(void *arg)
6028 struct rb_test_data *data = arg;
6030 while (!kthread_should_stop()) {
6031 rb_write_something(data, false);
6034 set_current_state(TASK_INTERRUPTIBLE);
6035 /* Now sleep between a min of 100-300us and a max of 1ms */
6036 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
6042 static __init void rb_ipi(void *ignore)
6044 struct rb_test_data *data;
6045 int cpu = smp_processor_id();
6047 data = &rb_data[cpu];
6048 rb_write_something(data, true);
6051 static __init int rb_hammer_test(void *arg)
6053 while (!kthread_should_stop()) {
6055 /* Send an IPI to all cpus to write data! */
6056 smp_call_function(rb_ipi, NULL, 1);
6057 /* No sleep, but for non preempt, let others run */
6064 static __init int test_ringbuffer(void)
6066 struct task_struct *rb_hammer;
6067 struct trace_buffer *buffer;
6071 if (security_locked_down(LOCKDOWN_TRACEFS)) {
6072 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
6076 pr_info("Running ring buffer tests...\n");
6078 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
6079 if (WARN_ON(!buffer))
6082 /* Disable buffer so that threads can't write to it yet */
6083 ring_buffer_record_off(buffer);
6085 for_each_online_cpu(cpu) {
6086 rb_data[cpu].buffer = buffer;
6087 rb_data[cpu].cpu = cpu;
6088 rb_data[cpu].cnt = cpu;
6089 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
6090 cpu, "rbtester/%u");
6091 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
6092 pr_cont("FAILED\n");
6093 ret = PTR_ERR(rb_threads[cpu]);
6098 /* Now create the rb hammer! */
6099 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
6100 if (WARN_ON(IS_ERR(rb_hammer))) {
6101 pr_cont("FAILED\n");
6102 ret = PTR_ERR(rb_hammer);
6106 ring_buffer_record_on(buffer);
6108 * Show buffer is enabled before setting rb_test_started.
6109 * Yes there's a small race window where events could be
6110 * dropped and the thread wont catch it. But when a ring
6111 * buffer gets enabled, there will always be some kind of
6112 * delay before other CPUs see it. Thus, we don't care about
6113 * those dropped events. We care about events dropped after
6114 * the threads see that the buffer is active.
6117 rb_test_started = true;
6119 set_current_state(TASK_INTERRUPTIBLE);
6120 /* Just run for 10 seconds */;
6121 schedule_timeout(10 * HZ);
6123 kthread_stop(rb_hammer);
6126 for_each_online_cpu(cpu) {
6127 if (!rb_threads[cpu])
6129 kthread_stop(rb_threads[cpu]);
6132 ring_buffer_free(buffer);
6137 pr_info("finished\n");
6138 for_each_online_cpu(cpu) {
6139 struct ring_buffer_event *event;
6140 struct rb_test_data *data = &rb_data[cpu];
6141 struct rb_item *item;
6142 unsigned long total_events;
6143 unsigned long total_dropped;
6144 unsigned long total_written;
6145 unsigned long total_alloc;
6146 unsigned long total_read = 0;
6147 unsigned long total_size = 0;
6148 unsigned long total_len = 0;
6149 unsigned long total_lost = 0;
6152 int small_event_size;
6156 total_events = data->events + data->events_nested;
6157 total_written = data->bytes_written + data->bytes_written_nested;
6158 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
6159 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
6161 big_event_size = data->max_size + data->max_size_nested;
6162 small_event_size = data->min_size + data->min_size_nested;
6164 pr_info("CPU %d:\n", cpu);
6165 pr_info(" events: %ld\n", total_events);
6166 pr_info(" dropped bytes: %ld\n", total_dropped);
6167 pr_info(" alloced bytes: %ld\n", total_alloc);
6168 pr_info(" written bytes: %ld\n", total_written);
6169 pr_info(" biggest event: %d\n", big_event_size);
6170 pr_info(" smallest event: %d\n", small_event_size);
6172 if (RB_WARN_ON(buffer, total_dropped))
6177 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
6179 item = ring_buffer_event_data(event);
6180 total_len += ring_buffer_event_length(event);
6181 total_size += item->size + sizeof(struct rb_item);
6182 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
6183 pr_info("FAILED!\n");
6184 pr_info("buffer had: %.*s\n", item->size, item->str);
6185 pr_info("expected: %.*s\n", item->size, rb_string);
6186 RB_WARN_ON(buffer, 1);
6197 pr_info(" read events: %ld\n", total_read);
6198 pr_info(" lost events: %ld\n", total_lost);
6199 pr_info(" total events: %ld\n", total_lost + total_read);
6200 pr_info(" recorded len bytes: %ld\n", total_len);
6201 pr_info(" recorded size bytes: %ld\n", total_size);
6203 pr_info(" With dropped events, record len and size may not match\n"
6204 " alloced and written from above\n");
6206 if (RB_WARN_ON(buffer, total_len != total_alloc ||
6207 total_size != total_written))
6210 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6216 pr_info("Ring buffer PASSED!\n");
6218 ring_buffer_free(buffer);
6222 late_initcall(test_ringbuffer);
6223 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */